Try Intent-System

intent-cli is built with intent-cli

Today its own development runs as a single seat: one conductor, a fresh subagent for each independent review, and a second review by another vendor's model. Every PR keeps those review records, so you can check the workflow in public.

As of 2026-09-22, 643 of the 891 merged pull requests in the intent-system repository are units cut and tracked by intent-cli.

Let it question one feature you're about to build

Open your coding agent in a repository you care about and paste this, with a real feature in place of <topic>:

Grill <topic> with intent-cli.
Keep asking me one question at a time until the intent is packet-ready.

What happens: your agent reads intent-cli's grill guide, studies your existing code and docs first, lists the open questions in dependency order, and asks you the most blocking one — one decision at a time, without making you repeat what the repository already says.

grill itself is read-only: it writes nothing and publishes nothing. To keep your answers as decision records, the host needs to be initialized (intent init) with a domain; an issue comes later, from a packet drafted from those records. Both are part of the next step.

Take one unit all the way to a merged PR

solo-conductor is the single-seat team mode. Your one thread — the conductor — plays architect, orchestrator, and builder. Reviews stay independent: for every review and re-review, the conductor starts a fresh subagent that did not write the change, and its verdict is recorded as an independent review.

Run the thread wherever you already work — the Claude desktop app, Claude Code in a terminal, or an Orca Run — and let it start its own reviewer subagents inside the same session. There is no transport, pane layout, or seat roster to set up. How 1 thread + subagents works →

Hand this to your agent

Set up intent-cli in this repository as a single-seat solo-conductor team.
Follow intent-cli's host initialization guidance (dry-run first) and ask me the decisions that are mine.
Before recording team_mode, confirm every intent-cli that reads this host is v0.32.0 or later.
Record team_mode=solo-conductor, confirm guide bootstrap reports solo-conductor-complete,
then run the first unit with intent-cli guide solo-conductor.
Start a fresh subagent for every review, and check with me before publishing an issue or merging.

What your agent will run

1. 01
   Initialize the host

   intent init, dry-run first and then --write — the same flow as every other setup.

2. 02
   Record the mode

   intent-cli team-mode set --domain <domain> --team <team> --mode solo-conductor --write

3. 03
   Confirm it

   intent-cli guide bootstrap --domain <domain> --team <team> reports solo-conductor-team-bootstrap in state solo-conductor-complete.

4. 04
   Run the ten-step loop for each unit

   intent-cli guide solo-conductor lays out the loop — design claim, packet, issue, implementation, PR, independent review, fixes, merge, closeout — and the rules that keep the reviewer independent.

Optional: a second opinion from another vendor

Declare the team under [[cross_runtime_review.teams]] in .intent-cli/config.toml. On the repositories that declaration lists, approving a PR then needs a verdict from a different runtime — Codex, Claude, Cursor, Copilot, or OpenCode — recorded against the PR's current head commit, and publishing an issue through publish-flow needs a design verdict on the packet. intent-cli renders the reviewer prompt and records the verdict; your seat runs the reviewer. This is how intent-cli itself is reviewed today.

These gates keep an honest workflow honest; they are not a security boundary. Other routes, such as gh pr merge, are not gated.

Limits: one unit at a time, and only on an agent runtime that can start isolated subagents — if yours cannot, use the multi-seat setup. Already running the conductor in an Orca Run? It can be recorded for the solo team as well (session-layer topology record-orca-run).

Sharing the host with other machines or seats? Upgrade every intent-cli that reads it to v0.32.0 before you record the mode. An older intent-cli rejects a team-mode.json that contains solo-conductor, and the whole file then fails to load for every team on that host.

Shape intent and publish an issue without writing code

With team_mode=authoring-only, you can shape intent and publish an issue without writing code. The agent interviews intent and authors packets; implementation may be handled by another team, a human, or a vendor. This mode has no delivery topology or delivery seats, and delivery supervision does not apply.

Use team_mode=authoring-only for <domain>.
Interview me about intent, author the packet, and ask for approval before publishing the issue.
Do not create delivery topology or seats; report the authoring audit and external handoff.

The released contract, in one place

Command-emitter boundary: intent-cli supplies deterministic guidance and records or verifies workflow state. The coding agent or operator runs herdr, providers, and any scheduler after the relevant decision; intent-cli does not launch them, operate herdr, or register a scheduler.