Launch: harness-foundry v0.4 — versioned agent harnesses with traces & human-gated evolution

[cobusgreyling]

Launch thread — harness-foundry

Official home for launch chatter, Show HN follow-ups, and first-run questions.

Composable harness runtime for production agents — primitives, traces, evolution, outerloop governance.

→ Repo · Showcase · Launch kit · QUICKSTART

---

The problem

Most production agents still run on harness mush: host-specific prompts (Cursor rules, Claude Code config, ad-hoc scripts). When something breaks, there is no shared vocabulary, no session trace, and no path from run → tune → govern.

What this is

Model  →  Primitives  →  Runtime  →  Trace  →  Evidence
         (compose)      (session)    (jsonl)   (outerloop)

harness-foundry is the missing runtime layer between model SDKs and reliable behaviour:

You get	Detail
Declarative stacks	stack.yaml from versioned primitives
Sessions + locks	foundry run → stack.lock + JSONL traces
Trace-driven evolution	L1 report → L2 proposal (human gate before apply)
Host adapters	Cursor / Claude Code — same stack, same traces
Governance seam	Optional outerloop evidence → verdict
CI	Reusable foundry-gate workflow

Not another LangGraph. Not another model SDK. Primitive-first harness engineering.

Ecosystem

loop-engineering  →  harness-foundry  →  outerloop
   (patterns)         (runtime)          (governance)

Repo	Role
loop-engineering	Design reliable inner loops
harness-foundry	Execute composable harness primitives
outerloop	Evidence, verdict, answerability

Try it in 60 seconds

npx @cobusgreyling/harness-foundry init --from minimal
foundry validate
foundry run --goal "Verify harness wiring"
foundry sessions list
foundry trace show --session <id>
foundry evolve report --session <id>

Or clone and dogfood:

git clone https://github.com/cobusgreyling/harness-foundry.git
cd harness-foundry && pnpm install && pnpm build && pnpm demo

+ IDE hosts: foundry host integrate cursor · foundry host integrate claude-code
+ Governance: examples/with-outerloop

Status

v0.4.0 — implementer execution, host adapters, npm publish.

Roadmap next: OpenAI adapter, real MCP stdio, foundry evolve apply (still human-gated).

→ ROADMAP · CHANGELOG

FAQ (common first questions)

Question	Short answer
vs LangGraph / ADK?	Graph orchestration vs primitive-first harnesses + traces + evolution
vs Cursor rules?	Host-portable, versioned stacks; same traces across Cursor / Claude Code / CI
Why not just prompts?	Prompts aren't locked, composed, traced, or evolved as a unit
Is this a framework?	Runtime + CLI for harness composition — bring your model / host
Auto-apply stacks?	No silent auto-tune. L1 report → L2 proposal → human review
Governance?	Optional outerloop seam for evidence → verdict → answerability

→ Full comparison: docs/vs-alternatives.md

How to use this thread

• First-run friction — paste errors, OS, command
• Show HN / social follow-ups — link external posts here so answers stay in one place
• Primitive ideas — or open an issue with the primitive request template
• Wins — share a stack.yaml or before/after that worked

Get involved

• Contributor start here
• Good first issues
• CONTRIBUTING

---

Own the runtime. Trace the runs. Govern the outcomes.

If you're posting Show HN / X / LinkedIn, ready-to-paste copy lives in docs/LAUNCH.md. Drop a link here when you ship so we can track Q&A in one thread.