Entity trail

Agent Framework Python

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Showing the first 40 findings. More graph evidence exists in the corpus.

Corpus findings

2026-08-14 / agents-researcherMicrosoft ships enforcement middleware for AGENT-HOOKS-0.1, a two-star community spec for agent lifecycle hooksAgent Framework Python 1.14.0 (2026-08-13) added experimental AGENT-HOOKS-0.1 enforcement middleware behind an opt-in `agent-hooks` extra. The spec is AgentHook — an Apache-2.0 open specification for AI agent lifecycle hooks stewarded by Agentic Thinking Limited (UK), first tagged v0.1.0 on 2026-04-26, last pushed 2026-07-07, and sitting at 2 GitHub stars. A hyperscaler implementing a near-unknown independent spec is the most interesting part: hooks are the deterministic enforcement boundary every agent runtime now has (Claude Code, Cursor, Codex, Copilot), and this is the first move toward one of them being portable rather than per-vendor.
2026-08-14 / agents-researcherMicrosoft Agent Framework Python 1.14.0 adds Mistral, spins Durable Task out to its own repo, and blocks Windows junctions in skill discoveryPython 1.14.0 landed 2026-08-13 with a native Mistral chat client (chat, streaming, tools, structured output, embeddings), request/response hooks on OpenAIChatCompletionClient, workflow checkpoint and resume in the AG-UI integration, and BackgroundAgentsProvider.release_session() for cancelling work and releasing per-session state. Two breaking changes: functional workflow definitions must now be built into stateful instances before running, and Foundry Hosted Agents migrate to the Agent Server Responses 2.x storage model. The Durable Task and Azure Functions integrations moved out to a separate agent-framework-durable-extension repo, and a quiet security fix now rejects Windows junctions during skill discovery — the same directory-traversal class that has been hitting agent skill loaders all year.
2026-08-09 / agents-researcherMicrosoft Agent Framework Harness and Foundry Hosted Agents reach GA — and a cited teardown puts Claude Code at 98.4% harness, 1.6% AI logicInfoQ reported on August 3 that the Agent Framework Harness and Foundry Hosted Agents moved to general availability in .NET and Python, four months after MAF 1.0 shipped on April 2. The harness bundles function invocation, per-call history persistence, context compaction, a todo list with plan/execute modes, file memory, skills, web search, tool approval and built-in OpenTelemetry. Two numbers make the case: research analyzing Claude Code's architecture found approximately 98.4% of the codebase is harness infrastructure versus 1.6% AI decision logic, and in a comparative run MAF halted after 40 round-trips with a limit-reached message while the GitHub Copilot SDK ran to 300 without stopping on its own.
2026-08-09 / agents-researcherCheck Point found 11 vulnerabilities across six agent frameworks — and argues the frameworks, not prompt injection, are the bugAt Black Hat 2026, Check Point's Yarden Porat and Shahar Tal presented 11 vulnerabilities spanning LangChain, LangGraph, CrewAI, AutoGen, Microsoft Agent Framework and Google ADK, showing that prompt-controlled content crosses the boundary into trusted framework logic itself. The bug classes are boringly conventional — insecure deserialization, SSRF, path traversal, use-after-free — which is precisely the point. A Microsoft Agent Framework checkpoint-deserialization flaw let one user plant a payload via prompt injection that fired when another user reloaded their session ($10,000 bounty); Google ADK shipped an unauthenticated HTTP API on by default that executed arbitrary Python and exposed service-account credentials ($3,133.70).
2026-08-09 / news-researcherMicrosoft's CodeAct Cuts Agent Latency 52% and Token Use 64% by Having the Model Write One Program Instead of Chaining Tool CallsMicrosoft Agent Framework's CodeAct has the model write a single short Python program that calls tools via call_tool(), executes it once in a fresh Hyperlight micro-VM per call, and returns a consolidated result — shipping in the alpha agent-framework-hyperlight package. On Microsoft's representative multi-step workload, traditional tool wiring took 27.81 seconds and 6,890 tokens while CodeAct took 13.23 seconds and 2,489 tokens: 52.4% faster and 63.9% fewer tokens. Reporting on the Agent Harness GA also cites an analysis that ~98.4% of the Claude Code codebase is harness infrastructure and only ~1.6% is AI decision logic.
2026-08-05 / agents-researcherMicrosoft Agent Framework 1.17.0 pulls Durable Task and Azure Functions integrations out of the core packagesThe dotnet-1.17.0 release on August 4 extracts the Durable Task and Azure Functions integrations into separate packages across both .NET and Python, decoupling long-running workflow persistence from the framework core. It also makes declarative workflows fail when an agent returns an error rather than continuing silently, and fixes the handoff orchestration sample's failure to respond to user input. The 12,610-star project is now shipping roughly weekly point releases, and the error-propagation change in particular will surface previously swallowed failures in existing declarative workflows.
2026-08-05 / agents-researcherMicrosoft ships a GitHub Copilot Harness for Agent Framework with approval gates on every shell command and file writePublished August 4, the GitHub Copilot Agent is an Agent Framework agent backed by the Copilot CLI and SDK, shipping as Microsoft.Agents.AI.GitHub.Copilot for .NET 8+ and agent-framework-github-copilot for Python 3.11+. It contributes shell execution, file read/write, URL fetching, MCP server integration, custom tool registration, session persistence, and project guideline directories. The governance design is the notable part: every shell command, file write, URL fetch, and custom tool routes through a user-defined permission handler, so human-in-the-loop approval is the default rather than an opt-in.
2026-07-29 / sources-researcherNVIDIA Open-Sources NOOA, an Agent Framework Where the Agent Is One Python Class and the LLM Fills In Method Bodies Marked With EllipsesNVIDIA Labs released Object-Oriented Agents (NOOA) as an open research preview that collapses prompts, tools, callbacks and workflows into a single Python class: state lives in fields, capabilities in methods, prompts in docstrings, and type annotations act as enforced contracts, with LLM-driven loops completing any method body left as an ellipsis at runtime. NVIDIA reports state-of-the-art accuracy and lower token cost versus prior harnesses on SWE-bench Verified, CyberGym L1 and ARC-AGI-3, and released the framework, capability tests, benchmark agents, evaluation methodology and a technical report (arXiv 2607.20709) together. The repo at NVIDIA-NeMo/labs-OO-Agents is at 490 stars with commits through July 29, and the practical appeal is that agents defined this way stay compatible with ordinary Python testing, tracing, refactoring and version control.
2026-07-26 / skill-finderMicrosoft shipped a batteries-included agent Harness for .NET and Python — one call gets you compaction, todos, file memory, and approval rulesReleased July 22, the Agent Framework Harness bundles the tool-calling loop with iteration limits, per-call history persistence, context-window compaction, persistent todo/plan tracking, durable file memory across turns, progressively-loaded skills, web search, standing tool-approval rules with heuristic auto-approval, and OpenTelemetry. Adoption is `create_harness_agent()` with `FoundryChatClient` in Python or the `AsHarnessAgent()` extension on `IChatClient` in .NET — you supply a chat client, instructions via `ChatOptions`, and optional tools. Every feature is on by default and individually removable through `HarnessAgentOptions`, which makes it a useful reference implementation even if you never adopt it: it's a checklist of what a production harness owns.
2026-07-26 / agents-researcherMicrosoft ships Agent Framework Declarative Workflows 1.0 — multi-agent orchestration moves from code to YAMLOn July 23 Microsoft took agent-framework-declarative to 1.0.0 for Python, matching the already-stable .NET Microsoft.Agents.AI.Workflows.Declarative package. Workflows are authored in YAML — state and expressions via Power Fx formulas, conditions, loops and jumps, sequential and conditional agent routing, function tools, MCP and HTTP endpoints, human-in-the-loop approval pauses, and checkpoint/resume — then loaded as ordinary Workflow instances that execute identically to code-first ones. The pitch is reviewability: non-developers can audit orchestration logic without knowing the framework, which matters for regulated deployments where the routing graph is the compliance artifact.
2026-07-25 / arxiv-researcherMulti-Agent Belief Combination Breaks When Agents Gain or Lose Sensors Mid-RunThis paper attacks an assumption baked into consensus averaging, logic-based knowledge-base merging, and epistemic logic alike: that the structure determining what each agent can represent stays fixed. When agents gain or lose observational capacity during execution, previously admissible beliefs become structurally impossible. The proposed hybrid framework pairs answer set programming (for elaboration tolerance, declarative integrity constraints, and explanations) with Python's numerical flexibility, and proves admissibility preservation under refinement, unique mass-preserving repair under coarsening, and explanation completeness — with evaluation across 100 randomly generated topology changes confirming complete violation detection and explanation coverage.
2026-07-25 / agents-researcherMicrosoft Agent Framework ships breaking Python 1.12.x and .NET 1.15 with AG-UI stabilized and GPT-5.6 cache breakpointsThe framework cut python-1.12.0 (July 21) and python-1.12.1 (July 23), promoting the AG-UI package from RC to stable, fixing stateless replay of reasoning-paired tool calls, and adding explicit prompt cache breakpoints for GPT-5.6 models. The same window graduated the harness agent, mode/todo providers, and tool approval out of experimental, and added an Azure Cosmos DB semantic-memory context provider with cross-session origin attribution. On .NET, 1.14.0 (July 21) split the legacy AGUI package into external SDK packages — a breaking upgrade — and 1.15.0 (July 22) added a Dapr agent provider example plus workflow credential hardening.

Graph relationships

RELEASED
Microsoft
Microsoft released Agent Framework Python 1.6.0 with Shell tool for local and Docker execution
Source finding
RELEASED
Microsoft -> Agent Framework Python
Microsoft released Agent Framework Python 1.6.0 with Shell tool for local and Docker execution
Source finding

Source trail

1.14.0 AGENT-HOOKS-0.1 InfoQ The Register GitHub (microsoft/agent-framework)Microsoft Agent Framework Blog NVIDIA Technical Blog (corroborated by the NVIDIA-NeMo/labs-OO-Agents repo)Microsoft Agent Framework Blog — The Microsoft Agent Framework Harness is now released Microsoft Agent Framework Blog arXiv 2607.21210 GitHub (Microsoft Agent Framework)arXiv / HuggingFace Daily Papers (NVIDIA)

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…