MetaHarness — mint a custom AI agent harness from any repo. Browser Studio + `npx metaharness` CLI. Runs on Claude Code, Codex, pi.dev, Hermes, OpenClaw, RVM.

Evaluation + Darwin-Mode bridge for @metaharness/workspace-lens: turn Jacobian-Lens interpretability receipts into a flywheel-consumable workspace_probe score, and reject structurally-brittle prompt/policy mutations (final answer up, workspace grip down).

Temporary demo launcher for MetaHarness — try `npx cognitum metaharness <command>`. Not the final package/name.

MetaHarness example — Claude Code workspace + plugin scaffold with CLAUDE.md, .claude/, and MCP wiring (npx @metaharness/claude-code my-bot)

MetaHarness example — multi-agent research dossier scaffold for Claude Code with scout, web search, grading, synthesis (npx @metaharness/research my-bot)

MetaHarness example — Hermes cli-config harness scaffold for AI agents (npx @metaharness/hermes my-bot)

MetaHarness example — Claude Code incident-response vertical with responder, runbook-runner, escalator, postmortem agents (npx @metaharness/devops my-bot)

MetaHarness example — bioinformatics harness with gene lookup, PubMed literature retrieval, and genomic sequence fetch via NCBI E-utilities and Ensembl REST (npx @metaharness/example-bio my-bot)

AI red-teaming for the AI agents & LLM apps you own: stress-test them with adversarial models to find security failures (prompt injection, tool misuse / excessive agency, data leakage, jailbreaks, denial-of-wallet), auto-patch (blue team), retest, and get

A verifiable self-improvement loop for agent harnesses. Freeze the model. Evolve the harness. Promote only what proves lift.

Cost-optimal model router for AI agent harnesses — route each query to the cheapest model that's good enough (k-NN over labelled embeddings). The productized DRACO Phase-2 finding.

Freeze the model, evolve the harness. Two measured applications: (1) SWE-bench code-repair — conformant GLM->Opus empty-patch cascade resolves 51.3% Lite (n=300) and 55.6% Verified (278/500, Wilson 95% CI [51.2, 59.9], official swebench gold eval, no gold

MetaHarness example — Hugging Face SDK showcase: model/dataset discovery, serverless inference, and Space exploration (npx @metaharness/example-huggingface my-bot)

MetaHarness example — Google Cloud Platform scaffold wired to Cloud Storage, BigQuery dry-run, and Vertex AI Gemini inference (npx @metaharness/example-gcp my-bot)

Fine-tune cheap open-source LLMs (GLM, Qwen, DeepSeek) on your AI coding agent's successful runs with LoRA (SFT + DPO) so your model cascade escalates to expensive frontier models (GPT, Claude) less often — cutting cost-per-resolved. Turns run history int

MetaHarness scaffold for Stripe billing agents — subscriptions, refunds, and webhook handling in test mode by default (npx @metaharness/example-stripe my-bot)

MetaHarness example — Microsoft Azure scaffold: resource management, Blob storage, and Azure OpenAI via DefaultAzureCredential (npx @metaharness/example-azure my-bot)

Cross-platform kernel for the agent-harness-generator — wasm primary, NAPI-RS native fallback

MetaHarness example — sales pipeline pod (qualifier, opener, closer) for Claude Code (npx @metaharness/sales my-bot)

MetaHarness scaffold for Blockchain / web3 — read-chain, event streaming, and simulate-tx on Ethereum Sepolia testnet via viem (npx @metaharness/example-web3 my-bot)