Field guide · GitHub data as of 22 Oct 2026

Open agent
frameworks

The open source projects people use to build and run AI agents, sorted by the job they do. Stars show attention, not quality. Licence and activity tell you more about whether a project is safe to build on.

Start here

Rising now

Stars gained in the four weeks from 21 September to 18 October 2026. Large projects lead on raw gain. Small, new projects lead on growth rate.

Most stars gained

    Fastest growth (1,000+ gained)

      Where the intelligence lives

      Every agent is a model inside a harness. An open, auditable stack that no single vendor controls needs as much of the work as possible to live in the harness, as files and code that people can read, version and run on any model. Each row in this guide shows which jobs its harness takes from the model.

      Stays with the model
      JudgmentDeciding the next step when the situation does not match a known pattern.Most model-centred here: UI-TARS (a vision model), Strands and smolagents (a thin loop by design).
      Can move into the harness

      What the evidence says

      Changing the harness did not change the average score. It changed results by task type, and it changed cost. One researcher ran two models on the same 80 private coding tasks. Each model ran in its maker's own harness (Claude Agent SDK, OpenAI Codex SDK) and in a neutral one (LangChain Deep Agents).

      • Average pass rates were close. Claude Opus 4.8 solved 48.8% in its own harness and 50.0% in the neutral one. GPT-5.5 solved 55.6% and 54.4%. Neither gap is statistically clear.
      • The averages hid a split. On repository tasks, Anthropic's harness did 9 points worse. On programming-contest tasks, it did 24 points better. The author chose this split after seeing the data, so it needs a second study.
      • The neutral harness used 1.2 to 1.6 times as much model spend per solved task.

      So in this study the harness mattered most for fit to the task and for cost, which an open harness can control. It is one single-author preprint, not yet peer reviewed.

      Source: Arjmandi, "Harness or Model?", arXiv 2609.11987, Sep 2026

      Saved skills can move ability from a strong model into the harness, but a bad skill can hurt.

      • One study turned successful agent runs into reusable procedures that the harness can run and check. On the ALFWorld and WebArena benchmarks, success went up and agent turns went down. A small open model (Qwen3.5-9B) that used procedures built by GPT-4o solved 98.5% of ALFWorld tasks. GPT-4o, the model that built them, solved 96.3%.
      • A second study found 307 cases where a loaded skill caused a failure (125) or made a run slower or more costly (182). Skills that looked relevant often caused the agent to leave out required work or add verification steps that were not needed.

      Hermes Agent and Stagehand build this idea in. Hermes writes skill files from experience, and Stagehand caches browser actions. Their own docs describe the features but do not measure them.

      Sources: Guo et al., "Skill-DisCo", arXiv 2606.26669, Jun 2026 · Dong et al., "Agent Skills Can Be Harmful", arXiv 2608.11888, Aug 2026

      Some builders keep the harness thin on purpose. AWS calls Strands "model-driven": the model directs its own steps and tool calls. Anthropic advises "finding the simplest solution possible, and only increasing complexity when needed." Both designs bet that better models will need less harness.

      Sources: AWS, "Introducing Strands Agents" · Anthropic, "Building effective agents"

      Some projects move work back into the model. Agent Lightning uses reinforcement learning to train the model inside an existing agent on that agent's own runs. The procedure the harness learned becomes model weights.

      Sources: Agent Lightning, arXiv 2508.03680 · v1.0 technical report, arXiv 2608.17528

      Researchers now treat the model and harness as one system. A June 2026 survey splits the harness into six jobs: observation, context, control, action, state and verification. It argues that agent quality comes from how the model, the harness, the task and the evaluation fit together.

      Source: Guo et al., "From Question Answering to Task Completion", arXiv 2606.20683

      A sketch of an open default stack

      One way to build agents that are secure, auditable, international and not tied to one company, using projects in this guide. This is a starting map for discussion, not a tested design.

      LayerOpen partsWhy it fits
      ModelOpen-weight models from several countries (DeepSeek, Qwen, Kimi and GLM from China, Mistral from France, Llama and gpt-oss from the US) behind a routerThe model becomes a part you can swap. No single lab can switch the stack off.
      ProtocolsMCP for tools, A2A for agent-to-agent callsBoth are now governed by Linux Foundation bodies, not by the companies that created them.
      Control and stateLangGraph or Microsoft Agent Framework (US), Haystack (Germany)Saved state lets a person pause, inspect, replay and resume every run.
      SkillsSkill folders in the open Agent Skills format, as Hermes Agent and the Claude Agent SDK useProcedures are plain files you can review, version and share between harnesses.
      MemoryMem0, Graphiti, Hindsight, LettaMemory lives in your own database, not in a vendor account.
      GuardNVIDIA OpenShell, E2B, one container per agent (as in NanoClaw)The harness, not the model, decides what files, network and secrets an agent can reach.
      AuditOpenTelemetry traces from the control layer, with cost records (as in Paperclip)Each action and each dollar can be traced to an agent, a person and a policy.
      Team layerQM, Paperclip, Onyx, n8n (Germany)Accounts, permissions and budgets for an organization, on infrastructure you run.