Agents inside a SaaS product
Specialist agents that use company context and tools, with human review, per-customer budgets and a run history behind every output.
Stablehand
Run agents inside your infrastructure, on any harness, and keep every run as learning signal.
Stablehand is the runtime at the centre of our stack. Each session runs in its own sandbox, the harness you choose drives the model, and every turn, tool call, approval, correction and dollar spent lands in your own Postgres. Swap the harness or the model; the record of how your work gets done stays with you.
- Status
- Early access
- Language
- TypeScript
- Stack
- Fastify, PostgreSQL, React
- Licence
- MIT
3agent harnesses on one control plane
17/17model and harness combinations certified across seven providers
2,430turns in a ten-wave endurance run, with zero failures
300/300benchmark tasks run and graded in one campaign
Why we built it
The best agent framework changes every quarter. The record of how your company does its work should outlast all of them. Every team we built agents for was rebuilding the same runtime underneath (sandboxes, approvals, streaming, budgets), and every hosted platform kept the most valuable output, the traces and corrections, in someone else’s cloud.
We built Stablehand to keep that record inside your infrastructure. A small core owns runs, events, budgets, approvals and artifacts, and everything else, from evals to domain tools, plugs in as an extension.
What it does
- Three harnesses on one control plane: Claude Agents, LangGraph and PI share tools, events, approvals, budgets and the operations UI
- One sandbox container per session, so a crash fails a single turn and the conversation carries on
- A durable event log: every turn, message, tool call, approval and feedback event is written to Postgres before it streams, with replay and signed webhooks
- Human review that survives restarts, with approval cards, editable documents and structured field requests
- Background work under governance: triggers, schedules, agent chaining and a durable turn queue with backpressure
- Cost accounting you can trust: budgets per agent and per team, and a model-gateway spend ledger
- Revocable, capability-scoped API keys for every extension and integration
- Docker Compose in minutes, and a Kubernetes driver that gives each session its own job
Where it’s going
- Evals where your agents run: rubrics as versioned artifacts, judges as agents and release gates as approval cards
- One campaign runner for eval suites, reinforcement-learning environments and pass@k, on your own cluster
- Learning signal you keep: full trace export in a training format, with corrections turned into eval cases automatically
- Promotion gates and a score ledger, so a new model or adapter takes over only when it wins on your evals
- Your own fine-tuned open-weight models, served on partner GPUs behind the same gateway
- More harnesses, Codex among them, and microVM isolation
Who it’s for
Back-office workflows
Claims intake, finance close and document review as chained agents, with approvals before anything is filed and the event log as the audit trail.
Model and harness bake-offs
Run the same private tasks across harness and model combinations, grade them, and route each task to the cheapest configuration that passes.
Work with us
Tell us what the work is, what data it touches and what constraints apply. We’ll work out what to build and how to measure it.