Ask once. Your coding agent builds the orchestration team in the background.
This page is for the multi-seat team, for when you need several units in flight at once. Most work runs well in one thread with subagents — see that page first. Keep the product conversation here, or give design its own herdr pane. The guide emits the setup and safety commands; your coding agent or operator runs them, records delivery topology, and verifies readiness. It pauses when a decision or permission needs you, while intent-cli remains a deterministic command emitter and state verifier.
Ask once. Watch a herdr-only team take shape.
This existing recording starts with three blank worker panes. A coding agent follows an intent-cli setup contract and turns them into a working orchestration team in about 10 minutes. It is a 47-second, 15× setup reference—not a fabricated v0.32.0 transcript.
Start this work in a herdr-only four-thread team for <owner>/<repo>. Use intent-cli's bootstrap guide, ask me for the CLI/model for each seat and the application kind, and emit commands only. After I approve, I will run herdr, provider, and scheduler commands; record and validate the topology, then report the first verified handoff.
- 01Design stays with you — here or in herdr
Keep product intent here, or give design its own herdr pane.
- 02Three worker roles appear in herdr
Orchestrator, implementation, and review start in dedicated panes and folders.
- 03Topology and liveness are verified
intent-cli records the supplied mapping, validates it, and the agent reports a ready handoff.
Who does what: your coding agent operates herdr and launches the chosen AI providers; intent-cli supplies and validates the deterministic workflow contract. Credentials, trust, or permission decisions still pause for you — they are never guessed in the background.
Design
Intent, packets, product decisions
Orchestrator
Verifies state, dispatches one ready unit
Implementation
Loopless receiver: implement one issue
Review
Loopless receiver: review against intent
herdr-only is recommended for collocated teams (fewer dependencies). agmsg + herdr is deprecated but still works and is the unrecorded default; neither transport is primary.
See one intent slice move through the team
A guided 30-second view of the operating contract — from a ready packet to verified closeout.
Make one intent packet ready
- Canonical input
- Intent tree + product decisions
- Team action
- Resolve meaning, priority, and acceptance boundaries.
- Verified output
- A reviewable ready packet
In intent-cli v0.32.0, herdr-only is the recommended transport for collocated teams because it has fewer dependencies. agmsg + herdr is deprecated but still works and remains the unrecorded default; neither transport is primary, and intent-cli and GitHub remain authoritative.
One visible team. One chosen session layer.
herdr-only session transport
Recommended for collocated teams because it has fewer dependencies; herdr makes dispatch, completion, and supervision visible.
Four threads, clear boundaries
Design can stay human-facing or use its own pane; herdr supervises the visible team.
intent-cli + GitHub
Deterministic guidance, topology validation, queue, issues, PRs, CI, and closeout remain authoritative.
agmsg + herdr
Deprecated in v0.32.0. It still works and remains the default when no transport is recorded, so record herdr-only explicitly when you move. Removal is planned only after current users are checked.
The released contract, in one place
Released baseline
intent-cli v0.32.0 is the current stable release. A team records one of three team modes: solo-conductor (one thread plus a fresh review subagent per review, the recommended start), delivery (a multi-seat team for several units in flight), or authoring-only. The three-thread timer loop remains a supported alternative.
Install channels
npm i -g intent-system installs the intent-cli command with no .NET SDK required; the .NET global tool is JTechJapan.IntentSystem.Cli. intent-cli update derives its channel from the executable path.
Cross-runtime review
For a team declared under [[cross_runtime_review.teams]], on the repositories it lists, PR approval needs a verdict from a different runtime (Codex, Claude, Cursor, Copilot, OpenCode) recorded against the current head commit, and publish-flow needs a design verdict on the packet. intent-cli renders the prompt and records the verdict; it never launches the reviewer. The gates are not a security boundary.
Acceptance and supervision
Guide primacy is an acceptance rule: a capability whose guide route is missing or wrong is not shipped. Supervision is opt-in per team ([supervision] opt_in_teams) and corroborates findings within the cycle before escalation.
Session choices
For collocated multi-seat teams, herdr-only is recommended because it has fewer dependencies. agmsg + herdr is deprecated: it still works and remains the default when nothing is recorded. Neither transport is primary. An Orca Run you already use can be recorded against a seat or a solo team.
File-backed delivery
Declare delivery_method: file-backed to persist a durable, addressable task envelope and deliver a one-line pointer. Without that declaration, delivery remains inline.
Named branch lanes
A lane comes from the registry/default or an explicit packet draft --lane choice. This is a released preview-through-1.x capability, not a 1.0 compatibility guarantee. The packet records branch_lane and routing_snapshot; intent-cli proposes and records the lane but does not create or manage branches.
Git-backed multi-user claims
claim acquire and claim verify use immutable Git-backed records under .intent-cli/claims/. This is a released preview-through-1.x capability, not a 1.0 compatibility guarantee. Only the successful plain push owns the claim, so claim before scaffold and verify ownership before work proceeds.
Prompt class and scope
Prompt policy uses literal prompt classes and scoped approval, including answerable_by, a hard non-overridable risk floor, and live-dialog compare-and-swap. This is a released preview-through-1.x capability, not a 1.0 compatibility guarantee. Unknown or unmatched cases escalate instead of guessing.
Authoring-only team mode
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 provides no delivery topology or delivery seats; delivery mode remains the default, and delivery supervision is not applicable in authoring-only mode. This is a released preview-through-1.x capability, not a 1.0 compatibility guarantee.
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.
How the four roles fit
Design owns meaning
Humans and AI shape intent, packets, priority, release scope, and decisions that require judgment.
The orchestrator owns flow
It reads canonical intent-cli and GitHub state, publishes ready work, delegates one unit, and verifies every claim.
Implementation and review stay independent
Loopless receivers work in dedicated folders. The reviewer checks the PR against packet and intent, not the implementer's local story.
Messages wake work; they do not define truth
In herdr-only mode, the coding agent or operator uses the recorded topology and herdr to deliver bounded tasks and route completion back. agmsg + herdr is deprecated but still works; GitHub and intent-cli remain authoritative.
Evaluate orchestration by operating guarantees
Agent count and parallel execution are easy to demo. These eight questions reveal whether a system can keep real work moving safely across sessions.
Persistent intent
Intent trees and packets preserve why the work exists across agent sessions.
Canonical source of truth
intent-cli and GitHub — not a pane, message, or agent claim — define workflow state.
Structural role separation
Design, orchestration, implementation, and review have separate authority and workspaces.
Verified liveness
Read the pane, confirm live attachment, and require a fresh end-to-end round trip.
Artifact-first completion
PRs, tests, CI, and running results prove completion instead of a done message.
Fail-closed behavior
Ambiguous ownership, readiness, or results return to design or the operator.
Post-merge write-back
Closeout updates canonical intent state before the next slice is selected.
Provider independence
intent-cli supplies deterministic guidance; it does not launch or choose an AI provider.
Choose herdr-only when the whole team lives in herdr
The four-role model and its authority boundaries do not change with transport. What changes is how a concrete task is delivered, observed, and reported.
| Path | Today | Delivery | Best for |
|---|---|---|---|
| herdr-only | v0.32.0 · released · recommended | intent-cli resolves logical role → recorded workspace and pane; herdr carries bounded tasks and completion reports. | Teams that want fewer moving parts, visible supervision, and multiple isolated teams on one machine. |
| herdr + agmsg | deprecated · still works · unrecorded default | herdr supervises the workspace; agmsg carries delegation signals and replies. | Existing agmsg teams until they move to herdr-only. |
Neither transport is primary, and the multi-seat team is not the only shape: one seat can run solo-conductor instead. v0.32.0 also lets you record adopted Orca Runs against seats (session-layer topology record-orca-run), require cross-runtime review per team, and opt teams into supervision. Choose one transport and one driver mode for a domain/repository, and keep intent-cli's command-emitter boundary intact. Read the v0.32.0 release notes →
Say “build it in herdr-only mode”
The guide supplies the canonical setup, safety gates, and topology-writing commands. Your coding agent or operator runs herdr and the chosen providers after you approve them; intent-cli does not launch providers or operate herdr. You can keep working in the design conversation while the agent reports a verified handoff or asks for a decision.
Start this work in a herdr-only four-thread team for <owner>/<repo>. Use intent-cli's bootstrap guide, ask me for the CLI/model for each seat and the application kind, and emit commands only. After I approve, I will run herdr, provider, and scheduler commands; record and validate the topology, then report the first verified handoff.
Safety boundary
- In a multi-seat team, exactly one driver mode applies per domain/repository: orchestrator-message or the timer-loop alternative. Never mix them. A single-seat solo-conductor team is a separate team shape, where one conductor seat runs every step of a unit in order.
- The orchestrator coordinates ready work; it does not invent intent, product direction, release scope, or semantic decisions.
- Receiver startup reports are claims, not proof. Read the pane, verify live attachment, then require a fresh end-to-end round trip.
- Ambiguous or unreadable state fails closed and returns to design or the operator instead of guessing.
Follow the current guide's supervision, timeout, and pause policy. Polling is an observation aid, not a substitute for a recorded claim or verified handoff.
Turn the model into a verified first team
Generate the setup checklist with your coding agent, inspect the architecture, or bring your workflow to J-Tech Japan for hands-on adoption support.