Forge

Forge is the execution control plane for agent-driven teams. One sentence in, shipped software out. It turns an objective into accountable work contracts, runs AI agents on them in sandboxes, and keeps people in charge of every call that matters.

Ship at agent speed.

Forge is in a private beta and only open by invitation. Ask for an invite if you run a team that would put it to work.

§The problem

Shipping software is slow, and most of that time isn't spent building. It's spent waiting: on standups, on reviews, on design, on legal, on whoever owns the next decision. Agents can do much of the work in minutes, but on their own they're a liability in production. Who decided this? What did they touch? Who checked it?

Human-driven on the left, agent-driven on the right. A still from the launch film, which is a dramatization.

§What Forge is

Forge is an execution control plane for cross-functional initiatives. You give it one objective, such as “ship offline sync to 120,000 users in 14 days.” It compiles that into a live graph of work contracts and runs them with AI agents and people together:

  • Work contracts, not tasks. Each contract states the outcome, the evidence that proves it's done, its dependencies, who or what executes it, what that executor is allowed to do, and who approves.
  • A live dependency graph. Contracts unblock each other the moment their prerequisites are proven. Everyone sees the same state as a board, a plan, a timeline, or a personal “needs your attention” queue.
  • Agents that run safely. Each agent run gets a fresh sandboxed workspace, scoped tools, a spend cap and no production credentials. Shell commands and anything consequential pause until a person approves them.
  • Decisions that propagate. Record a decision once, for example “Pro is $12/month”, and every contract it affects is flagged to recheck. Agents get recorded decisions and approved upstream work as authoritative context.
  • Done means proven. Completion needs evidence and an approval recorded after the latest submission. Whoever submitted the work can't approve it.
  • Everything on the record. Who decided, what changed, and which evidence proved it, for every artifact.

§How a contract moves

relaunchdraftreadyrunningin reviewcompletedunblocksdependents
Ember while the work is unproven, green once it is verified. An agent can pause a run to ask a question or request approval, and a failed run can be relaunched.

§Under the hood

  • Agent runtime: the Claude Agent SDK harness running Claude Opus 5.5, with OS sandboxing, path confinement, per-run spend caps, and human approval for shell actions. Run logs stream live to the UI.
  • API: Fastify (TypeScript). Session auth (scrypt, hashed opaque tokens), role-based access per organization and objective, a contract state machine, dependency propagation, and server-sent events for live updates.
  • Data: Postgres, with embedded PGlite for zero-setup local development.
  • Client: a React and Vite web app, plus an Electron desktop shell.
  • Shipping: a single Docker image serves the API and web app. Integration tests cover the API and execution engine, including tenant isolation.

§What's next

A separate execution worker so deploys never interrupt agents, a first design-partner launch, GitHub integration for pull requests and checks, then SSO.

The Forge mark with the line “Ship at agent speed.”
Request an inviteFollow along on X