microsoft--agent-framework/python/packages/core/AGENTS.md

Core Package (agent-framework-core)

The foundation package containing all core abstractions, types, and built-in OpenAI/Azure OpenAI support.

Module Structure

agent_framework/
├── __init__.py          # Lazy runtime public API exports
├── __init__.pyi         # Public API typing surface for lazy root exports
├── security.py          # Public security primitives, middleware, and tools
├── _agents.py           # Agent implementations
├── _clients.py          # Chat client base classes and protocols
├── _types.py            # Core types (Message, ChatResponse, Content, etc.)
├── _tools.py            # Tool definitions and function invocation
├── _middleware.py       # Middleware system for request/response interception
├── _sessions.py         # AgentSession and context provider abstractions
├── _skills.py           # Agent Skills system (models, executors, provider)
├── _mcp.py              # Model Context Protocol support
├── _telemetry.py        # User-Agent identity and internal feature-usage mask
├── _workflows/          # Workflow orchestration (sequential, concurrent, handoff, etc.)
├── openai/              # Built-in OpenAI client
├── azure/               # Lazy-loading entry point for Azure integrations
└── <provider>/          # Other lazy-loading provider folders

Core Classes

Root Public API (__init__.py / __init__.pyi)

• agent_framework.__init__ uses lazy module-level __getattr__ for most public exports to keep cold import agent_framework lightweight.
• Keep _LAZY_MODULE_EXPORTS, _LAZY_EXPORTS, the explicit runtime __all__, and __init__.pyi synchronized whenever adding, removing, or moving a root public export.
• Runtime __all__ is still required for from agent_framework import *; the .pyi file is for type checkers and editors and does not replace runtime exports.
• Public deprecation behavior for a lazy export belongs in the owning module. The root package should delegate via the normal lazy export map instead of carrying one-off branches.

Agents (_agents.py)

• SupportsAgentRun - Protocol defining the agent interface
• BaseAgent - Abstract base class for agents
• Agent - Main agent class wrapping a chat client with tools, instructions, and middleware

Chat Clients (_clients.py)

• SupportsChatGetResponse - Protocol for chat client implementations
• BaseChatClient - Abstract base class with middleware support; subclasses implement _inner_get_response() and _inner_get_streaming_response()

Types (_types.py)

• Message - Represents a chat message with role, content, and metadata
• ChatResponse - Response from a chat client containing messages and usage
• ChatResponseUpdate - Streaming response update
• AgentResponse / AgentResponseUpdate - Agent-level response wrappers
• Content - Base class for message content (text, function calls, images, etc.)
• ChatOptions - TypedDict for chat request options

Tools (_tools.py)

• ToolProtocol - Protocol for tool definitions
• FunctionTool - Wraps Python functions as tools with JSON schema generation
• @tool decorator - Converts functions to tools
• use_function_invocation() - Decorator to add automatic function calling to chat clients

Middleware (_middleware.py)

• AgentMiddleware - Intercepts agent run() calls
• ChatMiddleware - Intercepts chat client get_response() calls
• FunctionMiddleware - Intercepts function/tool invocations
• AgentContext / ChatContext / FunctionInvocationContext - Context objects passed through middleware. A tool can declare a FunctionInvocationContext parameter to receive it; context.tools is the live, mutable tools list for the run, and context.add_tools(...) / context.remove_tools(...) enable progressive tool exposure (changes apply on the next function-calling iteration).
• MessageInjectionMiddleware - Session-scoped chat middleware that lets tools or other code enqueue messages for the next model call in the current AgentSession; it drains queued messages into the next call and loops only when no function calls need to be handled by the function invocation layer.

Sessions (_sessions.py)

• AgentSession - Manages conversation state and session metadata
• SessionStore - Experimental in-memory opaque session_id -> AgentSession snapshot store; reads return independent copies
• FileSessionStore - Experimental msgspec file-backed session snapshot store with atomic last-writer-wins updates; JSON is the default, serialization_format="msgpack" enables binary MessagePack, opaque keys are encoded to portable filenames, and only syntactically malformed snapshots are quarantined (schema, version, and state-decoder failures preserve the original file)
• register_state_type - Registers custom AgentSession.state classes with stable, process-wide type IDs and optional mapping codecs. Provider modules own registration for their custom state types and should use package-qualified IDs. Implicit Pydantic registration remains temporarily with DeprecationWarning, but module-level registration is needed to guarantee cold-start restoration.
• ServiceSessionId - Mapping alias for structured service-owned continuation handles used in AgentSession.service_session_id
• SessionContext - Context object for session-scoped data during agent runs. extend_messages(...) can attach ordered, deduplicated origin_session_ids attribution when a provider injects content from other sessions.
• ContextProvider - Base class for context providers (RAG, memory systems)
• HistoryProvider - Base class for conversation history storage
• InMemoryHistoryProvider - Built-in session-state history provider for local runs
• FileHistoryProvider - Experimental append-only file-backed history provider; msgspec JSON Lines is the default and serialization_format="msgpack" uses length-prefixed binary MessagePack records. Custom dumps/loads remain as deprecated JSON-only compatibility hooks and emit DeprecationWarning when supplied.

Skills (_skills.py)

• Skill - Abstract base for a skill definition bundling instructions (content) with frontmatter metadata, resources, and scripts. Concrete subclasses (InlineSkill, FileSkill, ClassSkill) accept a frontmatter=SkillFrontmatter(...) argument carrying the spec fields. Adding new spec fields is done in one place — on SkillFrontmatter — keeping the subclass constructors stable.
• SkillFrontmatter - L1 discovery metadata for a skill (name, description, license, compatibility, allowed_tools, metadata). All fields are mutable plain attributes; the constructor validates name, description, and compatibility against the spec but post-construction assignments are not re-validated. Spec fields are reachable on every skill via skill.frontmatter.
• SkillResource - Named supplementary content attached to a skill; holds either static content or a dynamic function (sync or async). Exactly one must be provided.
• SkillScript - An executable script attached to a skill; holds either an inline function (code-defined, runs in-process) or a path to a file on disk (file-based, delegated to a runner). Exactly one must be provided.
• SkillScriptRunner - Protocol for file-based script execution. Any callable matching (skill, script, args) -> Any satisfies it. Code-defined scripts do not use a runner.
• SkillScriptArgumentParser - Public type alias for an optional callable (raw args: dict | list[str] | str | None) -> dict | None that converts the raw args value before an InlineSkillScript runs (applied before the inline list-args guard). It is an opt-in customization hook (port of .NET PR #6498) that lets callers support backends sending tool-call arguments in a non-conforming shape (e.g. vLLM JSON strings). The output is constrained to a dict (named keyword arguments) or None, because inline scripts bind arguments by keyword name. Supply it via the argument_parser= constructor arg on InlineSkillScript, InlineSkill (default for scripts added via @skill.script), or ClassSkill (default for scripts discovered via @ClassSkill.script). When None (the default), the raw value is used unchanged. File-based scripts are unaffected (their runner owns arg handling).
• SkillsProvider - Context provider (extends ContextProvider) that discovers file-based skills from SKILL.md files and/or accepts code-defined Skill instances. Follows progressive disclosure: advertise → load → read resources / run scripts. By default all three tools it exposes (load_skill, read_skill_resource, run_skill_script) are registered with approval_mode="always_require", so every skill operation needs approval. To run unattended, pass one of the static auto-approval rules to ToolApprovalMiddleware (via auto_approval_rules): SkillsProvider.read_only_tools_auto_approval_rule approves only the read-only tools (load_skill, read_skill_resource) while still prompting for run_skill_script, and SkillsProvider.all_tools_auto_approval_rule approves every skill tool including script execution. Both rules reject any call carrying a server_label so they stay scoped to this provider's local tools and never auto-approve a same-named hosted tool. Alternatively, for trusted skills, the constructor / from_paths kwargs disable_load_skill_approval, disable_read_skill_resource_approval, and disable_run_skill_script_approval (all default False) opt individual tools out of approval entirely by registering them with approval_mode="never_require" (the auto-approval rules only apply to tools that still require approval). The tool names are also exposed as class constants (LOAD_SKILL_TOOL_NAME, READ_SKILL_RESOURCE_TOOL_NAME, RUN_SKILL_SCRIPT_TOOL_NAME).
• FileSkillsSource - SkillsSource that discovers file-based skills by scanning configured root paths for SKILL.md. The configured root paths define the trust boundary and are used as given (a root may itself be a symlink); everything discovered below a root is link-checked and fails closed. _discover_skill_directories rejects any entry that is a symbolic link, junction, or other reparse point (via the shared agent_framework._filesystem.is_link_or_reparse_point helper) before descending into it, and rejects a directory whose SKILL.md is itself such a link — otherwise a link planted under a root would be adopted as the skill root, and since every later guard treats the skill root as the boundary and only inspects segments below it, the link itself would never be inspected. Resource and script discovery apply the same rule per path segment via _has_link_or_reparse_point_in_path. An OSError while inspecting an entry is treated as unsafe (skip / reject), never as "safe".
• MCPSkillsSource - SkillsSource that discovers Agent Skills served over MCP by reading the well-known skill://index.json (SEP-2640). Index entries are dispatched by their type (case-insensitive): skill-md entries become one MCPSkill each (its SKILL.md body and sibling resources are fetched on demand via resources/read), and archive entries are downloaded as a single ZIP / TAR / gzip-TAR blob and unpacked entirely in memory (via the private _ArchiveEntryLoader) into a FileSkill whose SKILL.md body drives it and whose sibling files (matching the resource extensions, within the search depth) become in-memory InlineSkillResource resources. Nothing is written to disk — there are no temporary directories to create, own, or prune (this is a deliberate divergence from .NET, which extracts archives to disk; it removes the temp-dir leak and the dangerous prune-of-unowned-subdirs footgun). Entries whose type has no handler (e.g. mcp-resource-template) are skipped. MCP-delivered scripts are never runnable: the loader emits no SkillScripts, so a bundled script can at most surface as a readable resource (and only if it matches the resource extensions — .py is not a default resource extension). The archive SKILL.md frontmatter name must match the advertised index-entry name or the skill is skipped. Extraction is hardened: a .. path-traversal ("zip-slip") member name raises via _normalize_archive_member_name and aborts the whole skill (like the file-count/size limits), non-regular TAR members (links/devices) are skipped, and file-count / uncompressed-size (_read_member_with_limit) / download-size limits are enforced. Archive behavior is configured with archive_* constructor kwargs (archive_resource_extensions, archive_resource_search_depth, archive_max_file_count, archive_max_size_bytes, archive_max_uncompressed_size_bytes) — Python uses plain kwargs, not a *Options object as in .NET. A non-"resource not found" error while downloading an archive propagates (so a failed CachingSkillsSource refresh does not overwrite a cached list with a partial result). Unlike .NET's AgentMcpSkillsSourceOptions.RefreshInterval, this source has no built-in refresh interval; wrap it in CachingSkillsSource(..., refresh_interval=...) for caching/refresh. This is a port of .NET PR #6631; the FileSkillsSource script_extensions/resource_extensions kwargs default to the built-in tuples and treat None as "use defaults" and an empty tuple as "discover none" (an empty tuple previously fell back to defaults). FoundryToolbox.as_skills_provider() forwards matching archive_* kwargs to this source.
• SkillsSource decorators - Skill sources are composable: SkillsSource is the abstract base, with concrete sources (InMemorySkillsSource, FileSkillsSource, MCPSkillsSource) and decorators that wrap an inner source — AggregatingSkillsSource (concatenate several sources), FilteringSkillsSource (predicate filter), DeduplicatingSkillsSource (first-wins by name), and CachingSkillsSource (cache the inner source's skills list). DelegatingSkillsSource is the abstract base for decorators. get_skills takes a SkillsSourceContext: every source/decorator implements async def get_skills(self, context: SkillsSourceContext) -> list[Skill] and forwards context to inner sources. SkillsSourceContext (frozen) carries the invoking agent (SupportsAgentRun) and optional session (AgentSession | None); SkillsProvider builds it from before_run's agent/session and passes it into the pipeline. FilteringSkillsSource's predicate is context-aware: Callable[[Skill, SkillsSourceContext], bool] (port of .NET #6797). Default caching is applied only to the built-in, context-independent leaf sources: for the Skill / sequence-of-skills / from_paths constructors, SkillsProvider builds DeduplicatingSkillsSource(CachingSkillsSource(<file|in-memory leaf>)) so expensive filesystem/network discovery runs once. A caller-supplied SkillsSource is used as-is — never auto-wrapped in caching or deduplication — because auto-caching a context-aware caller source in a single shared bucket would replay the first invocation's skills for later SkillsSourceContexts and leak skills across agents/tenants (matches .NET, whose custom-source constructor also adds no caching/dedup). Callers who want caching on a custom pipeline compose CachingSkillsSource(inner, cache_isolation_key_selector=...) themselves. disable_caching=True only affects the built-in leaf caching (it has no effect on a caller-supplied source, which is never cached). CachingSkillsSource shares a single in-flight fetch across concurrent callers (per cache key) and does not update its cache on a failed fetch, so the next call retries (an initial failure leaves the cache empty; a refresh failure keeps the previously cached list). By default all callers share one cache bucket; pass cache_isolation_key_selector=Callable[[SkillsSourceContext], str | None] to cache separately per key (e.g. per agent name) for context-aware inner sources — the key should be low-cardinality and stable, and returning None (or leaving the selector None) uses the shared bucket. By default a cached list never expires; pass refresh_interval=timedelta(...) (port of .NET CachingAgentSkillsSourceOptions.RefreshInterval) to treat a cached list as stale once it is older than the interval so the next call re-queries the inner source (useful when an inner source such as MCPSkillsSource changes over the process lifetime; a zero/negative interval makes every result immediately stale, and a failed refresh keeps the prior list and retries). Freshness is measured with a monotonic clock (time.monotonic()). SkillsProvider.__init__ / from_paths expose a cache_refresh_interval kwarg that is threaded into the built-in CachingSkillsSource (it has no effect on a caller-supplied source or when disable_caching=True). MCPSkillsSource and MCPSkill accept exactly one of client (a fixed ClientSession) or session_provider (Callable[[], ClientSession], resolved on every fetch); providing both/neither raises ValueError. Use session_provider when the underlying session may be swapped over time — e.g. a reconnecting MCPTool/FoundryToolbox whose session is replaced on reconnect — so cached MCPSkills keep fetching against the live session instead of a closed one (MCPSkillsSource forwards its provider to every MCPSkill it creates). A fixed client is safe only when the session outlives the skills.

Model Context Protocol (_mcp.py)

• MCPTool - Base wrapper that owns the MCP ClientSession and exposes the remote server's tools as FunctionTools.
• MCPStdioTool / MCPStreamableHTTPTool / MCPWebsocketTool - Transport-specific subclasses.
• Argument allowlist (_prepare_call_kwargs) - Before each tools/call, kwargs are filtered to an allowlist built from the tool's declared parameters (inputSchema.properties) plus any user-configured extras. The declared half comes from the server's advertised schema, and runtime kwargs (FunctionInvocationContext.kwargs, seeded from function_invocation_kwargs) are merged with the model-supplied arguments upstream in _call_tool_with_runtime_kwargs, so provenance is gone by the time the filter runs. A runtime kwarg is therefore forwarded whenever the server declares a property of that name, without the model mentioning it — the server, not the caller, decides which runtime kwarg names it receives. A tool that declares no usable properties (including schemas with additionalProperties: true) forwards only the configured extras. _MCP_FRAMEWORK_DENYLIST is a narrow safety net covering only non-serializable framework objects a server declares in its schema (those are dropped); it does not generalize to arbitrary caller-chosen names, and explicit extras always win. The reserved _meta key is never forwarded as an argument; trusted caller/runtime _meta is validated as MCP request metadata, model-supplied _meta is discarded in generated MCP functions, and metadata precedence is caller/runtime < OpenTelemetry < tools/list metadata.
• allowed_tools (constructor arg on all MCPTool subclasses) - Restricts exposed MCP tools by raw remote MCP tool identity. Prefixed local names remain accepted only when the raw remote name already matches its normalized form; normalized/local aliases do not authorize a different raw remote name. If multiple raw remote tool names map to the same local function name, tool loading raises ToolExecutionException instead of first-one-wins shadowing.
• Progressive MCP disclosure (use_progressive_disclosure, always_load) - When enabled on any MCPTool subclass, the initial model-facing surface is loader tools (list_mcp_tools / load_tool / unload_tool, prefixed by tool_name_prefix when configured) plus allowed tools selected by always_load and tools loaded earlier on the same MCPTool instance. list_mcp_tools only reports tools that pass allowed_tools; filtered tools are not listed or loadable. Loader tool names are reserved in progressive mode: remote MCP tools whose local generated name collides with a loader name are omitted from the initial/listed surface, and explicit load_tool calls return a model-visible message pointing callers to tool_name_prefix or excluding the colliding tool. load_tool accepts one tool name or a list of tool names and uses FunctionInvocationContext.add_tools(...) so the selected generated MCP FunctionTools become available on the next function-calling iteration while keeping existing approval mode, argument filtering, header-provider runtime kwargs, result parsing, OTel, and task behavior. unload_tool accepts one dynamically loaded tool name or a list of names and removes them from the live tool list and persisted progressive surface, but it does not remove tools configured in always_load. Invalid always_load entries are ignored like unmatched allowed_tools entries.
• additional_tool_argument_names (constructor arg on all MCPTool subclasses) - Opt extra argument names back into the allowlist. Accepts a Sequence[str] (applied to every tool) or a Mapping[str, Sequence[str]] keyed by remote tool name, where the reserved key "*" denotes global extras. It is configured only in user code at construction; there is no per-call/runtime override, so a model-issued tool call cannot change which names pass through — but note this constrains the model, not the server, which still widens the effective allowlist through its schema. To use a server that accepts additionalProperties: true, list the extra names here and then either (1) manually extend that tool's inputSchema (via the .functions list after connecting) so the model is prompted to supply them, or (2) supply the values yourself via function_invocation_kwargs. If a normal forwarded argument name is supplied by both the model and function_invocation_kwargs, the model-supplied value wins; _meta is the exception and only trusted runtime/caller metadata is used.
• function_invocation_kwargs and MCP servers - That dict is shared across every tool in the run, including every attached MCPTool, and any name in it reaches a server that declares a matching inputSchema property. header_provider does not mitigate this — it reads the kwargs without consuming them. To keep a credential out of tool arguments, source it outside function_invocation_kwargs: read a ContextVar inside the provider (this still allows a different value per request), configure a custom http_client, or use env for MCPStdioTool.
• Sampling guardrails (sampling_callback) - Passing client= advertises SamplingCapability so the server can send sampling/createMessage. Because remote servers are untrusted (confused-deputy risk), the default sampling_callback is deny-by-default and applies, in order: a per-session rate limit (sampling_max_requests, default _DEFAULT_SAMPLING_MAX_REQUESTS), an approval gate (sampling_approval_callback), and a maxTokens cap (sampling_max_tokens, default _DEFAULT_SAMPLING_MAX_TOKENS). The approval callback (constructor arg on all subclasses; exported type alias SamplingApprovalCallback) receives the raw CreateMessageRequestParams, may be sync or async, and must return truthy to approve. When it is None (the default) every sampling request is denied; pass lambda params: True to restore legacy auto-approve as an explicit opt-in. Requests and denials are logged at WARNING (content is not logged). The per-session counter resets in _reset_session_state.
• MCPTaskOptions (experimental, MCP_LONG_RUNNING_TASKS feature, frozen) - Per-tool-instance options controlling the SEP-2663 long-running task lifecycle. When the server advertises a tool with execution.taskSupport == "required", MCPTool.call_tool transparently routes through call_tool_as_task, which sends an augmented tools/call, polls tasks/get until terminal, and reinterprets tasks/result as a normal CallToolResult. Instances are immutable; replace via MCPTool.task_options = MCPTaskOptions(...). Fields:
  • default_ttl: timedelta | None — forwarded to the server as params.task.ttl (milliseconds). When None, the server's default applies.
  • cancel_remote_task_on_local_cancellation: bool = True — only gates the CancelledError path. Abandonment paths (see below) always cancel.
  • max_task_wait: timedelta | None — client-side deadline for the whole post-create lifecycle (poll + result fetch). When exceeded, raises ToolExecutionException and fires a best-effort tasks/cancel. None (default) means no client-side bound. Bounds sleeps, sends, AND reconnects via asyncio.wait_for.
• Permissive fallback: servers that ignore the augmentation (return CallToolResult directly) or reject the unknown task field with METHOD_NOT_FOUND / INVALID_PARAMS fall back to the plain session.call_tool(...) path so legacy servers keep working. An unparseable success response (server accepted the augmented call but returned a payload that is neither CreateTaskResult nor CallToolResult) does not fall back — it raises ToolExecutionException to avoid double-executing a side-effecting tool.
• Submit-vs-track reconnect policy: a dropped connection before a task_id is known raises ToolExecutionException("connection lost; task state unknown") without re-issuing the augmented tools/call, so a server that accepted the request but lost the response cannot be made to start the same operation twice; once a task_id exists, tasks/get / tasks/result reconnect once and retry against the same id (a shared _send_with_one_reconnect helper).
• Cancel-on-abandonment vs terminal failure: any path where the remote task may still be running (max-wait exceeded, hard McpError in poll, malformed tasks/get, second connection loss in poll/fetch, reconnect failure) fires best-effort tasks/cancel before raising. Terminal failures (failed/cancelled/input_required server-side, completed+isError, malformed tasks/result after server completed) do not cancel — the server is already done. _MCPTaskAbandoned is the private marker distinguishing the two.
• Transient poll retry: a slow tasks/get that surfaces as McpError(code=408 REQUEST_TIMEOUT) is retried (bounded by max_task_wait). All other non-connection McpErrors during poll are treated as abandonment. tasks/result does not get transient retry — the server has already completed, so a slow payload fetch is anomalous.

File Access Harness (_harness/_file_access.py)

• AgentFileStore - Abstract async store backing the file-access harness. Implementations expose write, read, delete, list_children, file_exists, search, and create_directory over forward-slash relative paths. list_children returns the direct children (files and subdirectories, subdirectories first) as FileStoreEntry instances; search accepts a keyword-only recursive flag (default False) and, when recursive=True, walks all descendants and returns file_name values relative to the search directory.
• InMemoryAgentFileStore - Dict-backed store suitable for tests and lightweight scenarios.
• FileSystemAgentFileStore - Disk-backed store rooted under a configurable directory. Enforces relative-path normalization, root containment, and rejects symlink/reparse-point segments to prevent escape.
• FileSearchResult / FileSearchMatch - SerializationMixin DTOs returned by search, carrying the matching file name, a context snippet, and the matching lines with 1-based line numbers.
• FileStoreEntry - SerializationMixin DTO returned by list_children, carrying an entry name and type ("file" or "directory").
• FileAccessProvider - ContextProvider that adds shared file-access tools (file_access_write, file_access_read, file_access_delete, file_access_ls, file_access_grep, file_access_replace, file_access_replace_lines) plus default usage instructions to each invocation. file_access_ls enumerates direct children (both files and subdirectories) as {name, type} entries with an optional glob_pattern, so the agent can walk the tree level by level; file_access_grep searches recursively from an optional base directory and returns relative file_name paths, scoped via an fnmatch glob_pattern (where * crosses /, e.g. *.md, reports/*). file_access_replace substitutes old_string with new_string (failing if not found, or if multiple matches and replace_all is false); file_access_replace_lines replaces whole 1-based lines with literal text (each new_line includes its own trailing newline; an empty new_line deletes the line, including its line break). All tools are registered with approval_mode="always_require" by default, so every file operation needs host approval. Pass disable_write_tools=True to advertise only the read-only tools. To run unattended you can disable approval at the source with disable_readonly_tool_approval=True (read, ls, grep) and/or disable_write_tool_approval=True (write, delete, replace, replace_lines), which register the affected tools with approval_mode="never_require"; alternatively, keep approval on and pass one of the static auto-approval rules to ToolApprovalMiddleware (via auto_approval_rules): FileAccessProvider.read_only_tools_auto_approval_rule approves only the read-only tools (read, ls, grep), while FileAccessProvider.all_tools_auto_approval_rule approves every file-access tool including the write tools. Both rules reject any call carrying a server_label so they stay scoped to this provider's local tools and never auto-approve a same-named hosted tool. The tool names are also exposed as class constants (WRITE_TOOL_NAME, READ_TOOL_NAME, DELETE_TOOL_NAME, LS_TOOL_NAME, GREP_TOOL_NAME, REPLACE_TOOL_NAME, REPLACE_LINES_TOOL_NAME). Unlike MemoryContextProvider, the store is intentionally shared across sessions and agents.

File Memory Harness (_harness/_file_memory.py)

• FileMemoryProvider - ContextProvider that gives an agent a session-scoped, file-based memory backed by the same AgentFileStore abstraction. Adds tools (file_memory_write, file_memory_read, file_memory_delete, file_memory_ls, file_memory_grep, file_memory_replace, file_memory_replace_lines) plus default usage instructions. Port of the .NET FileMemoryProvider.
• Scoping - Memories are isolated per session by default: each session writes under a working folder derived from context.session_id. Pass an explicit scope (e.g. a user id) to group memories across sessions, mirroring FoundryMemoryProvider's scope arg.
• Descriptions & index - file_memory_write accepts an optional description, stored in a companion <stem>_description.md sidecar. After each write/delete the provider rebuilds a capped (50-entry) memories.md index, and before_run injects that index as a user context message so the model knows what memories exist. Sidecars and the index are internal files hidden from file_memory_ls/file_memory_grep and rejected as write targets.
• DEFAULT_FILE_MEMORY_SOURCE_ID / DEFAULT_FILE_MEMORY_INSTRUCTIONS - Public defaults for the provider's source id and instruction banner.
• Harness wiring - create_harness_agent includes the FileMemoryProvider by default; the FileAccessProvider is opt-in and added only when a file_access_store is supplied (no implicit {cwd}/working store is created). Disable file memory via disable_file_memory; override its backing store via file_memory_store. When no file-memory store is supplied, the default is FileSystemAgentFileStore rooted at {cwd}/agent-file-memory. create_harness_agent also wires in MessageInjectionMiddleware by default (mirroring the .NET harness's UseMessageInjection); it is always on with no opt-out because it is a no-op when no messages are queued for the session.
• Experimental-feature gating - create_harness_agent itself is released (no longer @experimental), but a few features it can wire in remain experimental or pre-release: background agents (background_agents), file access (file_access_store), looping (loop_should_continue), and the shell tooling (shell_executor, from the pre-release agent-framework-tools package). Enabling any of them emits a single ExperimentalWarning (naming the responsible parameter). The three experimental harness providers share one HARNESS dedup key so the downstream experimental provider does not warn a second time; the shell tooling uses a separate dedup key (it is not a HARNESS-decorated feature).

Tool Approval Harness (_harness/_tool_approval.py)

• ToolApprovalMiddleware - Opt-in agent middleware that coordinates session-backed approval rules, heuristic auto_approval_rules, queued approval requests, collected approval responses, and streaming/non-streaming approval prompts. Heuristic callbacks receive the underlying function_call content.
• ToolApprovalRule / ToolApprovalState - Serializable state models for standing approvals and queued approval flow. ToolApprovalRule.arguments is None means a tool-wide rule; an empty dict {} means an exact no-argument call for create_always_approve_tool_with_arguments_response.
• create_always_approve_tool_response / create_always_approve_tool_with_arguments_response - Helpers that return normal function_approval_response content with additional_properties metadata consumed by ToolApprovalMiddleware. Standing rules for hosted tools include the server_label boundary, so same-named tools on different hosted servers do not share approvals.
• Mixed tool-call batches use a default .NET-style bypass in the function invocation loop: when a session is available, approval requests for known non-approval-required tools are treated as already approved, hidden, stored in session state keyed to the visible approval request ids from that batch, and reinjected only when that visible approval flow resumes.
• Approval resume is an immutable response boundary: the function invocation layer normalizes a private copy of caller messages, returns approved and rejected terminal results in the resumed response (and stream) before any final assistant message, and does not mutate the caller's approval Message or the earlier approval-request response.
• Approval/result correlation is occurrence-aware. A call_id may be reused after a completed round, so approval normalization matches ordered call occurrences and consumes approved results per occurrence rather than using one global result per call_id. All contents produced by one execution remain one result group and are consumed together, including multiple user-input requests.
• Approval resume keeps terminal function_result contents in tool-role messages and follow-up user-input requests in assistant-role messages, including mixed sibling batches.
• Function-call budget accounting counts one unit per executed result group, not per emitted function_result, so executions that pause for user input still consume max_function_calls.
• Once an invocation limit disables local tools, locally actionable calls and local approval requests are removed from streaming and final output. Provider-executed informational call/result pairs, hosted approval requests, and metadata-only stream updates remain visible.
• Declaration-only streamed calls emit their arguments only from the provider stream. The function layer sends a metadata-only follow-up (arguments=None) so id and user_input_request survive final aggregation without duplicating arguments.
• function_approval_request and function_approval_response are control-plane contents. History providers may retain them in their backing store for audit. The base HistoryProvider.before_run filters resolved wrappers from later model replay, but preserves unresolved requests/responses until a terminal result or follow-up request closes the occurrence. On unrelated turns the function layer omits the complete pending call batch from model input while retaining it for a later approval response. Providers configured with load_messages=False do not replay history, so this filter is intentionally not invoked.
• Reasoning content or opaque reasoning metadata bound to a function call is part of the same logical group as the call and terminal result. Function-loop replay and compaction must preserve that group atomically; adapters should fail before a stateless request when required reasoning cannot be reconstructed.
• Compaction scans assistant function calls and assistant- or tool-role terminal results in transcript/content order, pairing reused call_id values by ordered unambiguous occurrence. A result links only when exactly one preceding declaration remains unmatched; genuinely ambiguous duplicate declarations stay separate.

Agent Loop (_harness/_loop.py)

• AgentLoopMiddleware - AgentMiddleware that re-runs an agent in a loop by calling call_next() repeatedly (the pipeline re-reads context.messages each time). One configurable class covers two patterns: a required user should_continue predicate (sync or async, the first positional/keyword arg), and a chat-client judge built via the .with_judge(...) factory (a second chat client decides whether the original request was answered; loops while it is not, using a JudgeVerdict structured-output response — internally just an async should_continue predicate). The constructor covers the predicate pattern directly; only the judge has a convenience classmethod factory (.with_judge(judge_client, ...)) that forwards to __init__. Supports both streaming and non-streaming runs. By default a non-streaming run returns an aggregated AgentResponse containing every iteration's messages plus the injected next_message "nudge" messages (as user messages); set return_final_only=True to return only the last iteration's response. Streaming runs always yield each iteration's updates and emit the injected nudge messages as user updates between iterations (the return_final_only flag has no effect on streaming, and the final response reflects the last iteration; MiddlewareTermination is handled cleanly). should_continue is required; other constructor args are optional: max_iterations (safety cap; defaults to DEFAULT_MAX_ITERATIONS=10, explicit None→unbounded, positive int caps; .with_judge uses DEFAULT_JUDGE_MAX_ITERATIONS=5 as its default), next_message (defaults to a short "continue" nudge), return_final_only, and additional_instructions (an extra system message injected ahead of the input before the agent runs — becomes part of the original messages so it survives fresh_context resets and persists via a session). The judge is configured only through .with_judge (judge_client/instructions/criteria), not the constructor, and its reasoning is fed back to the agent as the next iteration's input; the judge forwards the original request messages and the agent's latest response messages verbatim so multi-modal content is preserved. criteria (a list[str]) is both injected as the agent's additional_instructions and rendered into the judge instructions wherever the {{criteria}} placeholder (CRITERIA_PLACEHOLDER) appears (DEFAULT_JUDGE_INSTRUCTIONS ends with it; custom instructions may include it, and it is stripped when no criteria are given). The should_continue/next_message callables are invoked with keyword args (iteration, last_result, messages, original_messages, session, agent, progress, feedback) and may be sync or async; declare only what you need plus **kwargs. should_continue may return a plain bool or a (bool, str | None) tuple whose second item is feedback surfaced to next_message/record_feedback via the feedback kwarg (the judge uses this to relay its reasoning). Stop precedence per iteration is max_iterations → should_continue, evaluated before record_feedback so the feedback is available to it.
  • Feedback tracking - record_feedback captures a per-iteration progress entry (called with the loop kwargs; if it returns a truthy string the entry is appended, otherwise the agent's response text is used as the fallback entry). The accumulated log is exposed to every callback via the progress keyword (a per-iteration copy of prior entries) and, when inject_progress=True (default), injected into the next iteration's input as a user message (the full log without a session, only the latest entry with a session to avoid duplicating history). fresh_context=True restarts each iteration from the original task plus the progress log; when a session is attached it is snapshotted (to_dict()) before the loop and restored (from_dict + field copy) between iterations so the local transcript and any service-side conversation id reset too (in-loop working-state is discarded, pre-loop state preserved, continuity carried only by the progress log).
• todos_remaining(*, looping_modes=None) / todos_remaining_message - Helper factories for todo-driven loops (the Python counterpart of .NET's TodoCompletionLoopEvaluator), designed for create_harness_agent but usable with any agent that registers a TodoProvider via context_providers. They resolve the TodoProvider/AgentModeProvider from the running agent (agent.context_providers, via _resolve_context_provider) rather than taking the provider as an argument, so they can be wired directly into loop_should_continue/loop_next_message. todos_remaining returns a should_continue predicate that loops while any todo is open; pass looping_modes=[...] to gate looping to specific operating modes (case-insensitive; honors the AgentModeProvider's source_id/available_modes), looping_modes=None (default) applies in every mode, and an empty sequence raises ValueError. todos_remaining_message is a next_message callable that lists the still-open todo titles and tells the agent to finish them, returning None when the session/agent/provider is unavailable or nothing is open (in which case the middleware's default None handling applies: reuse the previous iteration's messages verbatim under the default fresh_context=False, or DEFAULT_NEXT_MESSAGE only when fresh_context=True).
• background_tasks_running() / background_tasks_running_message - Helper factories for background-agent-driven loops, mirroring the todos_remaining pair. They resolve the BackgroundAgentsProvider from the running agent (agent.context_providers, via _resolve_context_provider) rather than taking the provider as an argument, so they can be wired directly into create_harness_agent's loop_should_continue/loop_next_message. background_tasks_running returns a should_continue predicate that loops while the provider's persisted state shows any task with status == RUNNING (pair it with max_iterations so the loop is bounded even if a task's persisted status is never refreshed). background_tasks_running_message is a next_message callable that lists the still-running tasks (#<id> (<agent_name>): <description>) and tells the agent to wait for them to finish and retrieve their results, returning None when the session/agent/provider is unavailable or no task is running.
  • Approval escape hatch - _has_pending_approval_request(result) checks whether an iteration's response carries a pending tool-approval request (any content with type == "function_approval_request"). Both the streaming and non-streaming loops stop and return that response to the caller before evaluating should_continue/max_iterations or injecting next_message, so the loop is HITL-safe even when wrapped outermost around a ToolApprovalMiddleware (mirrors the C# LoopAgent's HasPendingApprovalRequests).
  • Harness integration - create_harness_agent enables the loop when a loop_should_continue callable is passed; it prepends AgentLoopMiddleware(loop_should_continue, max_iterations=loop_max_iterations, next_message=loop_next_message) ahead of ToolApprovalMiddleware so the loop is the outermost middleware (each iteration is a full agent run including tool approval, and the escape hatch hands pending approvals back to the caller). loop_next_message and loop_max_iterations only take effect together with loop_should_continue (with no loop_should_continue there is no loop, so they are ignored); loop_max_iterations defaults to the loop's default cap (None → unbounded).

Workflows (_workflows/)

• Workflow - Graph-based workflow definition
• WorkflowBuilder - Fluent API for building workflows, including explicit output_from / intermediate_output_from selection for caller-facing emissions. output_from is an allow-list for Workflow Output; unselected executor payloads are hidden unless intermediate_output_from selects them as Intermediate Output. Use output_from="all" for explicit all-output behavior and intermediate_output_from="all_other" for visible progress from every output-capable executor not selected by output_from.
• WorkflowRunResult - Non-streaming workflow result with Workflow Output get_outputs() and Intermediate Output get_intermediate_outputs() accessors
• Functional workflow definition/build lifecycle - @workflow returns a stateless FunctionalWorkflowDefinition. Call build() to create a stateful FunctionalWorkflow scoped to one logical caller or session. The definition has no run() or as_agent() surface, so module-level decorated definitions cannot accidentally retain caller state. Each built workflow and its FunctionalWorkflowAgent must remain scoped to that caller/session. Pass a caller-scoped checkpoint storage to build(checkpoint_storage=...) when needed; hosts remain responsible for authorizing and tenant-scoping access to any shared checkpoint adapter.
• Orchestrators: SequentialOrchestrator, ConcurrentOrchestrator, GroupChatOrchestrator, MagenticOrchestrator, HandoffOrchestrator

Built-in Providers

OpenAI (openai/)

• OpenAIChatClient - Chat client for the OpenAI Responses API
• OpenAIChatCompletionClient - Chat client for the OpenAI Chat Completions API

Foundry (foundry/)

• FoundryChatClient - Chat client for Microsoft Foundry project endpoints
• FoundryAgentSessionStore - Experimental Foundry-hosting session store, lazily re-exported from agent-framework-foundry-hosting; currently backed by azure.ai.agentserver.core.storage.FoundryStateStore

Key Patterns

Creating an Agent

from agent_framework import Agent
from agent_framework.openai import OpenAIChatClient

agent = Agent(
    client=OpenAIChatClient(),
    instructions="You are helpful.",
    tools=[my_function],
)
response = await agent.run("Hello")

Using as_agent() Shorthand

agent = OpenAIChatClient().as_agent(
    name="Assistant",
    instructions="You are helpful.",
)

Middleware Pipeline

from agent_framework import Agent, AgentMiddleware, AgentContext


class LoggingMiddleware(AgentMiddleware):
    async def process(self, context: AgentContext, call_next) -> None:
        print(f"Input: {context.messages}")
        await call_next()
        print(f"Output: {context.result}")


agent = Agent(..., middleware=[LoggingMiddleware()])

Custom Chat Client

from agent_framework import BaseChatClient, ChatResponse, Message


class MyClient(BaseChatClient):
    async def _inner_get_response(self, *, messages, options, **kwargs) -> ChatResponse:
        # Call your LLM here
        return ChatResponse(messages=[Message(role="assistant", contents=["Hi!"])])

    async def _inner_get_streaming_response(self, *, messages, options, **kwargs):
        yield ChatResponseUpdate(...)