Skip to content
Conductor

The agents

Twelve narrow agents, three graphs, and a person at every gate

Each agent does one job with one structured model call, or no model at all. Routing is plain code. Where the stakes are high, the graph pauses and waits for you.

Hover a seat to lift it; click to open its card below. Blue cones plan, ink cylinders oversee, the amber ring communicates, the green stone distils.

Twelve agents, one narrow job each

Planning supervisor

One LangGraph graph with a deterministic router and three interrupts

Profiles, plan, task-list gate, assignment, assignment gate, provisioning. Change orders re-enter the same graph with a diff of added, modified and cancelled tasks.

Turns a résumé into a structured developer skill profile with one structured LLM call and no tools.

Reads
Résumé text (PDF or text upload)
Produces
Skills, seniority, stack familiarity per developer
LLM call: yesNo pause

profile.built

Decomposes BRD and FRD documents into modules and tasks with acceptance criteria, estimates, priorities and dependencies, plus a narrative plan.

Reads
Project documents; Repository state; Developer pools
Produces
Task library (TASK-1, TASK-2, ...); Implementation narrative
LLM call: yesNo pause

plan.created

The task-list gate: creates one approval, pauses the graph, and applies the decision exactly once when an admin resolves it.

Reads
The proposed task list
Produces
Approved backlog, or every task of this run cancelled
LLM call: noPauses for: The task list, before anything is provisioned

plan.approved, plan.rejected

Proposes who does what from the skill profiles, the project's developer pools and a configurable active-task capacity, then pauses for approval.

Reads
Approved tasks; Skill profiles; Live active-task counts
Produces
One assignment proposal per task
LLM call: yesPauses for: Every proposed assignment

assignment.proposed, assignment.approved, assignment.rejected

Creates one repository per approved spec, invites the assigned developers, installs the push and pull-request webhook. Best-effort and fully audited; it never raises.

Reads
Approved assignments; GitHub token
Produces
Repositories, collaborator invites, webhooks
LLM call: noNo pause

github.repo_created, github.webhook_installed, github.provision_failed

Reads the existing plan and a client change request and produces a validated task delta. It applies nothing.

Reads
Current plan; Change request text and documents
Produces
Added, modified and cancelled tasks
LLM call: yesNo pause

change.planned

The task-change gate: the delta waits for an admin before it is materialised into the backlog.

Reads
The proposed delta
Produces
Applied delta, or nothing
LLM call: noPauses for: The change delta

change.approved, change.rejected

Oversight pipeline

A non-interrupting graph: commit monitor, code review, completion

Autonomous but overridable. One LLM judgment (the review) sits between two pure functions (linking and completion), and every outcome is audited.

Links a pull request to its task by TASK id, floors the task's status to in review, and flags force-pushes and oversized pushes. Database only, no LLM.

Reads
A verified webhook event
Produces
Task link; Status floor; Anomaly audit rows
LLM call: noNo pause

anomaly.force_push, anomaly.oversized_push

The one LLM judgment in oversight: reads the real diff, the task's acceptance criteria and the active playbooks for the stacks touched, and records a verdict with severity-tagged findings.

Reads
PR diff; Acceptance criteria; Retrieved playbooks and past reviews
Produces
approve, comment or request changes; Findings; Acceptance met flag
LLM call: yesNo pause

code_review.created, code_review.fetch_failed

Closes a task as a pure function of real merge state and the persisted review. A merged PR that fails a gate is recorded as blocked with the reasons.

Reads
Linked PR state; Latest review for the merged head
Produces
Task complete (completed by agent), or completion blocked
LLM call: noNo pause

task.auto_completed, task.completion_blocked

Communication graph

Compose, optional approval gate, send

Feedback, alerts and the weekly executive digest are composed from real records, then sent or paused for approval depending on the configured policy.

Composes feedback, risk alerts and the weekly digest from real records and retrieved memory, then sends, or pauses for approval when the kind is gated.

Reads
Review verdicts, risk findings, digest numbers; Memory retrieval
Produces
Outbound email, threaded and embedded
LLM call: yesPauses for: Feedback and alert emails (the digest auto-sends)

email.approved, email.sent, email.send_failed

Standards

A single distillation pass, no graph and no interrupt

The Recon agent turns a document into a structured playbook; an admin activates it, and from then on code review retrieves it for every matching pull request.

Distils an uploaded coding-standards document into a structured rule set per technology stack. Embedding happens only when an admin activates the playbook.

Reads
Standards document; Stack slug
Produces
Distilled playbook (pending until activated)
LLM call: yesPauses for: Activation, by an admin

playbook.distilled, playbook.activated

What triggers each graph

Planning supervisor
An admin uploads documents and résumés, or files a change order
Oversight pipeline
Every signed GitHub webhook event
Communication graph
A review verdict, a risk finding, or the weekly schedule
Standards
An admin uploads a coding-standards document

See the agents run on a real repository

A walkthrough of a live tenant shows a plan approved, a pull request reviewed and a task closed by the rule.