Heros Agent

LLM Agentic Workflow Optimization Platform

[](LICENSE)

Point it at a codebase; it discovers the LLM call graph, and **opens verified pull requests that optimize your prompts, models, context strategies, and node wiring** — with statistical proof the change is better or cheaper before you merge.

Think **"Dependabot for LLM cost & quality."** You review a diff and merge; the platform never touches production behavior without evidence.

**Status: 22 of 34 phases are deployed and folded into the spec set; 12 still carry open work.** The four core subsystems and all three delivery surfaces are built: Discovery, the source-transformation Config Layer, the sandboxed Runtime, the Eval Harness, the attribution/diagnosis/verification engine, the autonomous optimizer, billing & metering, the admin and customer consoles, the offline `heros` CLI, and CI-mediated forge delivery all live under [`internal/`](internal/) and are demoable end-to-end through the `cmd/pNhermes` walkthroughs. All six optimization axes (P13 → P18) ship, each one **refusing by name** what its build cannot carry. The whole platform **stands up from one digest-pinned image set** on Docker Compose or Kubernetes ([`deploy/`](deploy/), see [P19](docs/prd/P19-deployment-delivery.md)), and it runs at [heros-agent.space](https://heros-agent.space).

**What is not finished is named, not implied.** Twelve phases keep an open change under [`openspec/changes/`](openspec/changes/) — the residue is acceptance runs and commercial cutover, not unbuilt subsystems: P21's live-mode Stripe cutover, P22's real-IdP integration, P19/P20's fresh-machine install matrices, and P28's first-owner path, which **does not work as shipped** ([`p28-first-owner-reachability`](openspec/changes/p28-first-owner-reachability/) is the follow-up and has