Agent Harness

An agent harness is the opinionated wrapper that turns a model into a working operator: context, tools, permissions, memory, verification, state, UI, and recovery around the model loop.

Why this page exists

The wiki already has strong concrete harness pages: Cursor Harness, Claude Code Harness, Codex Harness (OpenAI), Hermes Harness (Nous Research), OpenClaw Harness, and AI SDK HarnessAgent. It also has Harness Engineering, which describes the discipline. The missing primitive was the thing those pages all instantiate: the agent harness as a product and systems boundary.

This matters because 2026 agent capability is rarely explained by model quality alone. OpenAI's 2026 harness-engineering writeup describes engineers building a large internal product by steering agents, instrumenting feedback loops, and making the codebase and UI legible to Codex. Anthropic's agent guidance similarly emphasizes simple composable workflows, evaluator loops, routing, and orchestrator-worker patterns over magical autonomous blobs. The frontier is the model plus the environment around it. Source: OpenAI, https://openai.com/index/harness-engineering/, 2026-02-11; Anthropic, https://www.anthropic.com/engineering/building-effective-agents, 2026-06-18

Boundary stack

The harness is not just the runtime. The runtime executes. The harness decides what the agent is allowed to know, do, remember, verify, ask, and show.

What a harness must own

IRL comparison

The pattern is converging in the real world: agent runtimes are becoming swappable, and the harness layer is where product differentiation moves. Vercel's HarnessAgent makes this explicit by exposing Claude Code, Codex, Pi, and future harnesses behind one app-facing API. Source: AI SDK docs, https://ai-sdk.dev/v7/docs/ai-sdk-harnesses/overview, 2026-06-18; Vercel changelog, https://vercel.com/changelog/program-agent-harnesses-with-ai-sdk, 2026-06-12

Capsule routes

Runtime-specific knowledge should now route through Agent Harness Capsules before falling back to generic agent pages:

Generic harness questions still start here. Runtime-specific work loads the relevant capsule so the agent sees the right docs, bookmarks, tools, skills, and failure modes together.

2026 state of the art

• Harness portability. Applications want to swap agent harnesses the way they swap model providers. AI SDK HarnessAgent is the clearest current sign.
• Filesystem-first agents. Skills, instructions, schedules, traces, and artifacts increasingly live as files, not only hidden chat state. This validates Agent-Docs Mesh and Kevin's wiki-backed operating model.
• Markdown as substrate, not agency. .md files do not make an agent by themselves. The harness becomes interesting when agents use those files to select context, take actions, write updates, preserve state, and verify results. Source: X/@kevskgs, 2026-03-19
• Durable sessions. Agents need resumable workspaces, long-running state, and pause/resume semantics, not only stateless API calls. See Durable Agent Execution Architecture.
• Trace-first debugging. A serious harness captures the trajectory: model calls, tool calls, handoffs, guardrails, approvals, and artifacts. See Agent Trajectory Evaluation.
• Registry-governed capabilities. Tools, MCP servers, skills, CLIs, and remote agents need discovery plus local approval. See Agent Capability Registry.
• Agent legibility. Repos, logs, docs, and UIs should be readable by agents, not only humans. See Agent Legibility.
• Stack-wide source maps. Hitchhiker's Guide to Agentic AI is useful as a checklist for whether a harness design has covered the whole stack: model substrate, reasoning/alignment, retrieval and memory, harness/context management, skills/tools, MCP/A2A coordination, multi-agent topology, UI, evaluation, and deployment. It is a research map, not an implementation route by itself. Source: arXiv:2606.24937, checked 2026-07-04
• Training-environment standardization. Open-source models need shared environments for harness-aware RL instead of every trainer, harness, and benchmark rebuilding bespoke glue. OpenEnv is the current project to watch here: it standardizes a reset() / step() / state() environment loop while letting MCP serve as the agent-facing tool boundary for compatible environments. Source: Hugging Face, https://huggingface.co/blog/openenv-agentic-rl, 2026-06-08; Hugging Face docs, https://huggingface.co/docs/openenv/index, accessed 2026-06-20
• Local/open-weight harness stacks. A coding harness can be paired with a local model served through Ollama, llama.cpp, LM Studio, or an OpenAI-compatible runtime. That route optimizes for privacy, cost, modifiability, and learning, but it should be evaluated with the same receipt suite as hosted Codex/Claude Code runs. See Local Coding Agents. Source: Sebastian Raschka, 2026-06-27

Kevin rule

When creating or updating a concrete harness page, answer these five questions:

1. Who owns the loop?
2. Where does state persist?
3. How are capabilities discovered and permissioned?
4. What proves the agent is done?
5. How does a human inspect, interrupt, resume, or replay the run?

If a page cannot answer those, it is probably a tool note, not a harness article.

---

Timeline

• 2026-07-12 | Reabsorbed the spatial-context, Cloud/Desktop Hermes, Vercel SDK, and agent-skill-library launch cluster. It reinforces the harness boundary: agent capability is a product of the durable workspace, skill projection, human-visible surface, and verification path—not a model-only feature. Source: X Bookmark Ingest 2026-07-12
• 2026-07-04 | Added Claude Tag as the shared team-channel harness surface: Claude moves from private chat into Slack channels with scoped tools, channel memory, future planning, and ambient follow-up. Source: Anthropic, 2026-06-23; Source: X/@karpathy, 2026-06-23
• 2026-07-04 | Added Hitchhiker's Guide to Agentic AI as a stack-wide research map for harness-adjacent design coverage, while keeping concrete implementation routes on the specific harness/runtime/tool pages. Source: arXiv:2606.24937
• 2026-07-04 | Added Kevin's .md files don't make an agent row as the concise harness boundary: files are substrate, while agency is the loop that reads, writes, acts, and verifies through them. Source: X/@kevskgs, 2026-03-19
• 2026-06-29 | Added runtime-specific capsule routing via Agent Harness Capsules so Claude Code, Hermes, OpenClaw, and Codex each expose their own source-backed brain interface instead of sharing one generic agent bucket. Source: User request, 2026-06-29
• 2026-06-30 | Added Local Coding Agents as the open-weight/local runtime route for harness work, with receipt-based comparison against hosted harnesses. Source: X/@rasbt, 2026-06-27
• 2026-06-20 | Added OpenEnv as the open-source agentic RL environment standard to watch, because it sits below harnesses but may become the common training/evaluation socket for harness-aware open models. Source: User request, 2026-06-20; Hugging Face, https://huggingface.co/blog/openenv-agentic-rl
• 2026-06-18 | Created after a graph audit found many concrete harness articles but no canonical agent-harness concept page. Source: whole-wiki graph audit, 2026-06-18