Meta-Harness is a framework that automatically searches for task-specific harnesses, the surrounding code that controls what a fixed model stores, retrieves, and displays during work. Researchers and developers use it to optimize agent workflows, with reference experiments for text classification and terminal tasks; catalogue skills are examples of applying this approach.
Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx skills add stanford-iris-lab/meta-harness --skill meta-harness-terminal-bench-2git clone --depth 1 https://github.com/stanford-iris-lab/meta-harnessWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/stanford-iris-lab/meta-harness/meta-harness-terminal-bench-2)<a href="https://agentmods.dev/skills/stanford-iris-lab/meta-harness/meta-harness-terminal-bench-2"><img src="https://agentmods.dev/badge/skills/stanford-iris-lab/meta-harness/meta-harness-terminal-bench-2/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/stanford-iris-lab/meta-harness/meta-harness-terminal-bench-2"><img src="https://agentmods.dev/badge/skills/stanford-iris-lab/meta-harness/meta-harness-terminal-bench-2.svg" alt="Reviewed on agentmods" width="80" height="20"></a>- NVIDIA SkillSpector pass
What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00023 | $0.01662 |
| Opus 5 | $0.00012 | $0.00831 |
| Sonnet 5 | $0.00005 | $0.00332 |
| Haiku 4.5 | $0.00002 | $0.00166 |
Grade A, and why
meta-harness-terminal-bench-2 scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 12d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 130 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Meta-Harness (Terminal-Bench 2)
Run ONE iteration of agent scaffold evolution.
You do NOT run benchmarks. You analyze results + failed trajectories, propose agent variants, and implement them. The outer loop (meta_harness.py) handles benchmarking.
CRITICAL CONSTRAINTS
- You MUST produce 1 new agent variant every iteration
- Do NOT write "the frontier is optimal" or "stop iterating", or abort early.
Anti-overfitting rules
- No task-specific hints. Do not hardcode knowledge about specific tasks. Agents must be general-purpose.
- Never mention task names in agent code, prompts, or comments. No references like "if task contains 'async'" or "for polyglot tasks." If your improvement only helps one task, it's too specific.
- General guidance is OK. Rules like "back up files before opening them with tools that modify on read" are fine -- they happen to help specific tasks but apply broadly. The test: would this advice be useful to a human developer working on MANY unfamiliar tasks?
- If in doubt, make it more general. "Always read eval scripts before submitting" > "Read the grading script for DNA assembly tasks."
CONTEXT
You are evolving the AgentHarness agent scaffold for Terminal-Bench 2. It is located in agents/baseline_kira.py.
The search space is arbitrary Python code. You can override any method, call any library, make raw API calls, add new tools, change how the LLM is called, rewrite command execution, intercept and transform observations -- anything that's expressible in Python is fair game. The only constraint is that the agent must subclass harbor.agents.terminus_2.terminus_2.Terminus2 in the same way as baseline_kira.py does (for compatibility with the eval harness).
Model: Claude Opus 4.6. Evaluation uses the full official 89-task TB2 dataset. meta_harness.py chooses the trial count; the default path uses 2 trials per task.
Key files to read:
agents/baseline_kira.py- the full baseline implementation. Read to understand overridable methods.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 12d ago First seen · 130 lines · 23 tokens per session scan A 37d148f046ac
meta-harness-terminal-bench-2 is a skill published in the GitHub repository stanford-iris-lab/meta-harness (1,549 stars, last pushed today), licensed MIT. It adds 23 tokens to every session and 1,662 once invoked, about $0.0001 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.
Other skills, from other repositories
building
Implementation skill for writing production code with TDD. Covers the RED-GREEN-REFACTOR cycle, false-RED detection, vertical slicing, scope escalation, test process discipline, and code generation patterns. Loaded by component-builder and bug-investigator.
diff-driven-docs
Use when a BUILD phase completes, a commit is staged, or a PR is about to be created, and the diff has not yet been reflected in documentation. Also use when the user says "update docs", "sync docs", "document this", or asks whether documentation is up to date.
exploration
Two-mode exploration skill: (1) design dialogue — turn rough ideas into validated designs through collaborative interview before planning; (2) spike — throwaway code answering ONE design question, deleted or absorbed, never shipped. Router invokes mode via dispatch context.
memory-and-handoff
Two-mode skill: (1) session memory — load/persist durable workflow state under .cc10x/ (activeContext, patterns, progress) so context survives compaction; (2) handoff package — portable, secrets-redacted export for a coworker, different tool, or fresh non-cc10x session.
plan-review-gate
Use after saving a non-trivial plan or decision RFC when a fail-closed feasibility, completeness, and alignment review must block execution.
architecture
Greenfield architecture design: map functionality flows, draw components, design APIs, classify dependencies, plan observability. For multi-component, API, schema, auth, or integration-heavy work. For retrofitting existing code, use codebase-hygiene instead.