eve
A self-hosted agent platform that survives a hard kill — and resumes the human approval it was waiting on.
- React 19
- Node 24
- Postgres
- LiteLLM
- native Anthropic
- Tauri / Rust
Seven agents — a root orchestrator plus migration, PR-review, Sentry-fix and outreach specialists — run on a Postgres-backed workflow world. A run paused mid-flight on a human-approval gate survives a hard SIGKILL and resumes on restart, with idempotent side effects. A live 38-tool control plane sets approval policy per tool; a React 19 observability console sits on top. No third party in the data path — LiteLLM routes to native Anthropic, native OpenAI and a self-made Codex-pool gateway.
The problem
Most “AI agents” are a demo: a prompt loop with no memory of what it was doing when the process died, and no safe place to pause for a human. The moment you put one near production work — a migration, a PR, a Sentry fix — you need it to survive a crash mid-task and to stop and ask before it writes anything irreversible.
eve is the durable layer that makes that real: agents run as resumable workflows, not fire-and-forget scripts.
How it works
A root orchestrator delegates to specialists — code migration, PR review, Sentry-fix, outreach — over a Postgres-backed workflow world (graphile-worker). State is checkpointed, so a run paused on a human-approval gate survives a hard SIGKILL and resumes exactly where it stopped, with idempotent side effects so nothing double-fires on replay.
A live 38-tool control plane sets approval policy per tool (never / once / always) and lets an operator switch the model — LiteLLM to native Anthropic, native OpenAI, or a self-made Codex-pool gateway — with no restart. No third party sits in the data path.
What it proves
This is the strongest “agents in production, governed” evidence in the portfolio: durability, human-in-the-loop by design, and a self-hosted data path — the exact properties an enterprise needs before it lets an agent touch its code. Built at ADF; a React 19 observability console makes every run, tool call and approval legible.