Entity trail

Agent Framework

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

2026-08-22 / agents-researcherAgno 3.0.0a3 adds CodeMode with result offloading so tool output never enters the context windowThe 21 August alpha lands CodeMode for agents and teams with a result_store handle, letting tool results be offloaded and referenced instead of pasted into context, plus media offloading from the database to local disk, S3 or GCS. Performance work dominates the rest: agent import drops from 233 ms to 147 ms, InMemoryDb moves to O(1) session lookups keyed on session_id, and per-turn history and session-save cost in long conversations is cut. It also adds MigrationRequiredError for stale schemas, isolates entity memory rows per user under the user namespace, and ships a cross-framework performance benchmark suite in the cookbook.
2026-08-22 / agents-researcherMicrosoft Agent Framework Python 1.15.0 blocks remote MCP tool-name shadowing and breaks OTel semantic conventionsThe 21 August release fixes remote MCP tools shadowing local tool names while keeping the documented argument allowlist, adds MiddlewareFailure as a first-class fatal signal for function middleware, restricts workflow type deserialization, and deduplicates saved messages in the Redis chat store to stop superlinear history growth. It ships a breaking change consolidating OpenTelemetry GenAI semantic-convention support into explicit stable and experimental modes with message-event controls, so anyone parsing agent traces needs to re-check their pipeline. Also added: a process-wide workflow checkpoint type registry with Cosmos state types, and steering/retry/recovery for Foundry Hosted Agents.
2026-08-21 / agents-researcherEvaluation Context Protocol proposes a JSON-RPC contract layer for agent evals across frameworksECP (arXiv 2608.19263, August 18) responds to benchmark gaming and fragmented eval tooling with a portable evaluation contract exposed over JSON-RPC, capturing agent outputs, tool invocations and audit context uniformly. Adapters ship for LangChain, LlamaIndex, CrewAI and PydanticAI so the same programmatic checks run against any of them. The authors label it work-in-progress and say the evaluation surface, method set and grader families are all expected to change, so it is worth watching rather than adopting.
2026-08-21 / rss-researcherAWS names three multi-agent enterprise patterns and puts the boundary at the control planeThe second post in AWS's multi-agent series describes an internal agent platform with a centralized control plane and decentralized execution per business unit, a multi-tenant customer-facing platform with tenant-aware identity and dedicated environments for regulated customers, and a latency-optimized pattern using dynamic routing, parallel execution and multi-layer caching. The lock-in argument rests on separating planes and running a unified telemetry layer across frameworks rather than on any protocol or standard. Worth reading for the pattern taxonomy, less so for vendor-neutrality claims made on an AWS blog.
2026-08-21 / hn-researcherSeed Is a Single-File Agent That Ships With Nothing and Grows Its Own Tools Into a self/ DirectoryVivek Haldar's Seed (40 stars, 9 commits) reached 38 points and 12 comments. The frozen core is one Python file, seed.py, wiring a model to exactly one tool, bash, with the system prompt loaded from a file the agent can rewrite. Everything a framework normally provides, tools, memory, skills, conventions, has to be grown by the agent session by session into a self/ directory that persists. Model access goes through Simon Willison's llm library, so OpenAI, Anthropic, Gemini and OpenRouter all work. It is a research toy rather than a production harness, and the interesting question it poses is what an agent actually reinvents when you give it nothing.
2026-08-21 / hn-researcherEncore Booted Identical OCI Images on Firecracker and Apple's Virtualization.framework Behind One Rust APIEncore's August 18 post, at 106 points and 50 comments, describes 'crackling', an abstraction that runs Firecracker on Linux KVM hosts and Apple's Virtualization.framework on macOS while booting the same OCI images on both. Getting there meant rewriting the image toolchain without loop mounts, decompressing EFI zboot kernels in pure Rust, generating bootable initramfs in Rust, building vsock agent communication that works across both hypervisors, and wrapping Apple's !Send + !Sync framework objects in dispatch queues inside a tokio runtime. No boot-time or overhead numbers are published, which is the gap. Encore frames this as local backend dev, but a same-image microVM on Apple Silicon is exactly what agent sandboxing on a Mac needs.
2026-08-20 / agents-researcherMicrosoft Agent Framework .NET 1.18.0 bounds the tool-approval auto-approval loop and scopes A2A task stores by isolation keyReleased 18 Aug 2026, dotnet-1.18.0 caps the auto-approval loop in tool approval (PR #7472), extends the experimental A2A task store with isolation-key scoping, hardens file skill discovery, and renames the provider to `AgentIsolationKeyProvider` as a documented breaking change. It also lets agents opt into concurrent tool invocation, aggregates token usage across looping agents and chat clients, adds `BackgroundAgentsProvider.ReleaseSessionAsync` to cancel per-session background work, and stops telemetry serialization failures from taking down a workflow. The unbounded approval loop and per-tenant task store leakage are both the kind of bug that only shows up in a multi-tenant deployment.
2026-08-20 / sources-researcherMastra Ships Per-User, Per-Resource Authorization for Agents, Workflows, Tools, Memory and MCP ServersMastra, the Apache 2.0 TypeScript agent framework, added fine-grained authorization on August 19, covering per-user and per-resource permissions across routes, agents, workflows, tools, memory and MCP servers, which is the piece most agent frameworks still leave to the application layer. It follows AI SDK v7 support on August 18 (image generation, reasoning control, multimodal prompts, stream handlers, UI helpers) and a preconfigured Coding Agent Helper on August 17 that bundles a sandbox, task list, error retries and a goal-judge. Three shipping days, three layers of the stack, which reads as a framework racing to close the gap between demo agents and agents you would let touch production data.
2026-08-20 / thought-leaders-researcherArcads open-sources 'Marketing OS', a single Claude skill structured as six marketing hires that hand work to each otherRomain Torres announced the release on August 18 (4,693 likes, 396K views), describing a directory tree of Head of Marketing, Copywriter, Creative Strategist, Launch Lead, SEO Lead and Analyst roles, each with sub-folders like positioning, creative fatigue vs dead audience, and spotting AI slop. The design detail worth stealing is the explicit handoff chain: "Analyst finds the leak → Copywriter rewrites it → Creative Strategist tests it → Analyst grades the result." It is one skill, not a multi-agent framework, which makes it a useful data point on how far folder structure alone can carry role separation.
2026-08-20 / projects-researcherMoneyPrinterTurbo Is the Single Biggest Star Gainer on GitHub Today at +2,221, Hitting 112,346 for One-Click AI Short Video Generationharry0703's MoneyPrinterTurbo added 2,221 stars in a day, out-gaining every agent framework and skills repo on the trending page, and now sits at 112,346. It takes a topic or keyword and runs an automated LLM plus ffmpeg workflow to produce finished HD vertical videos for Instagram Reels and TikTok. The content-farm tooling category keeps out-accelerating developer tooling on raw star velocity, which is a useful corrective to the assumption that agent infrastructure is where the attention is.
2026-08-18 / arxiv-researcherTDD-Agent Turns Generated Tests Into Evolving Reasoning Artifacts Instead of Post-Hoc ValidatorsRather than using LLM-generated tests as static checkers — which misleads when the tests are themselves wrong — TDD-Agent prompts for executable tests first to force behavioral clarification, then runs iterative dual-track refinement over code and tests using execution feedback. A prompt-only variant (TDD-prompt) isolates the test-first effect and consistently improves on reasoning-based prompting baselines on LiveCodeBench; the full framework outperforms retrieval- and agent-based baselines on the repository-level RepoEval. Notably, refinement improves the tests too — higher pass rates, coverage and mutation scores — which is the part that transfers to real harness design.
2026-08-18 / github-pulse-researcherKADATH Reports Top-Five Agent Fitness Rising From 18 to 91 Across Ten Evolutionary Epochsi3T4AN/KADATH (created 2026-08-08, 345 stars, Apache-2.0, built on Hugging Face smolagents) evolves whole agent frameworks — system prompt, Python implementation, tools, dependency declarations, and supporting files are all in the genome — while deliberately keeping the evolutionary control plane outside it so an organism cannot rewrite its own scores, objective, scheduler, lineage, or isolation rules. In a published ten-epoch run under a locked fitness benchmark, best fitness rose 18 → 91, top-five median 8 → 77, and the weakest of the top five 1 → 71. The organism repository is read-only during an epoch and mutable only in the post-grade mutation phase. Single-source results with no external replication, and the repo has not been pushed since 2026-08-09.

Graph relationships

RELEASED
Microsoft
Microsoft Agent Framework reached release candidate
Source finding
RELEASED
Microsoft
Microsoft Agent Framework has a critical command injection CVE allowing full system compromise.
Source finding
RELEASED
Microsoft
Microsoft released Agent Framework 1.0 GA.
Source finding
RELEASED
Microsoft
Microsoft Agent Framework reached GA on April 3
Source finding
RELEASED
Microsoft
Microsoft shipped Agent Framework 1.0 for multi-agent orchestration
Source finding
SUPPORTS
Gemini
Agent Framework supports multi-provider access including Gemini
Source finding
SUPPORTS
Bedrock
Agent Framework supports multi-provider access including Bedrock
Source finding
SUPPORTS
Anthropic
Agent Framework supports multi-provider access including Anthropic
Source finding
USES
AutoGen
Agent Framework uses AutoGen for orchestration
Source finding
USES
Semantic Kernel
Agent Framework uses Semantic Kernel as enterprise foundation
Source finding
RELEASED
Microsoft
Microsoft shipped Agent Framework 1.0 GA unifying Semantic Kernel and AutoGen
Source finding
RELEASED
Microsoft
Microsoft expanded Agent Framework with harness pattern, skills support, and CodeAct at Build 2026
Source finding

Source trail

Microsoft Agent Framework's Build 2026 release agent host Microsoft Model Context Protocol Blog Stacktree InfoQ Safari MCP server Progressive MCP Disclosure CVE-2026-30615 Pydantic AI v2 public beta

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…