How Conductor works
From a requirements document to a merged, reviewed, closed task
Conductor's intelligence comes from how narrow agents are wired together: one structured model call per node, plain code between nodes, and a person at every gate that matters.
The lifecycle
- 01
Plan
Documents and résumés in; modules, tasks, criteria, estimates and dependencies out.
- 02
Gate
The task list waits in Decisions. Nothing is provisioned before you say so.
- 03
Assign
Owners proposed from skill profiles, developer pools and a configurable active-task capacity, then gated again.
- 04
Provision
Repositories, collaborator invites and webhooks created best-effort and audited.
- 05
Oversee
Every signed webhook event runs commit monitoring, code review and completion.
- 06
Complete
One deterministic rule closes a task, or records why it could not.
- 07
Communicate
Feedback, alerts and the weekly digest composed from real records, gated by kind.
Three graphs, one lifecycle
Narrow agents, deterministic control flow, explicit gates
Each agent is a node that makes one structured model call and returns data. Routing between nodes is plain code. The gates are where a person decides.
Admin-triggered and human-gated. A pure router with three interrupts; every database write lives in a node that runs exactly once.
- Profiles
- Planner
- Task-list gate
- Assignment
- Assignment gate
- Provisioning
Blue: an agent acts. Red: a person decides. Grey: a pure function.
One deterministic rule
A task closes only when every check passes. Anything else is recorded as blocked, with the reasons, and an admin can force-complete or reopen with a stated reason that lands in the audit log.
- 1A linked pull request is merged into the repository's default branch
- 2The latest review covers that exact merged commit
- 3The review did not request changes and found nothing high or critical
- 4The acceptance criteria are marked met
Blocked reasons: no_merged_pr, stale_review, request_changes, high_severity_finding, acceptance_unmet, task_blocked, unverified_head
- A linked pull request is merged into the repository's default branch
- The latest review covers that exact merged commit
- The review did not request changes and found nothing high or critical
- The acceptance criteria are marked met
Memory that retrieves rather than retrains
Emails, reviews, completions and activated playbooks are embedded into PostgreSQL with pgvector and retrieved before an agent composes or judges. Training on your standards means retrieval, never fine-tuning.
Packaged as one isolated tenant
Each subscription gets its own PostgreSQL database and encryption keys, its own API and console containers, its own GitHub credentials and mailbox, and its own model endpoint. Nothing is pooled between customers. Deploy anywhere Docker runs.
Risks and digests with no model in the loop
Risk monitor Daily
Computes delivery risks from the database with no LLM in the loop. Only the resulting alert email is written by a model, and repeat findings are latched for a configurable cooldown.
- Overdue tasks
- Stale in-progress tasks
- Blocked tasks
- Developers over capacity
- Assigned tasks with no commits
- Long-open pull requests
- Milestones at risk of their target date
Weekly digest Weekly
Rolls up real task, pull-request and review activity per project into the executive digest. The Communication agent only narrates the numbers.
- Tasks completed and by whom
- Pull requests merged and reviewed
- Open risks
- Decisions waiting
- BlockedTASK-151 Offline sync for field reports, blocked 6 days
- MilestoneMeridian field app, target date in 5 days with 8 open tasks
- CapacityTomas Ruiz has 3 active tasks (capacity 3)
- Stale PRPR #79 open 9 days without a merge
Repeat findings are latched for 7 days; only new ones raise an alert.
See Conductor on your own repositories
Book a walkthrough of a live tenant, or request a pilot on your GitHub organisation.