Summary
Add first-class agenting support to Devlane so teams can create specialized agents, assign work items to them, and optionally let those agents execute supervised autonomous tasks through existing Devlane and GitHub workflows.
Current state
Devlane already has useful foundations:
- Issue/work item assignment, labels, projects, cycles, modules, comments, and activity feeds.
- Instance-level AI settings for OpenAI credentials.
- GitHub App integration with workspace installation, project repository sync, PR-to-issue linking, push/PR/comment webhooks, and PR-driven issue state transitions.
- RabbitMQ-backed queue infrastructure for background tasks.
What is missing is the product and backend layer that treats agents as first-class actors with task lifecycle, permissions, GitHub execution, and human review.
Proposed capability
Users should be able to:
- Create workspace or project agents such as Bug Triage, Spec Breaker, PR Reviewer, Test Fixer, Docs Writer, Release Notes, and Coding Agent.
- Configure each agent's role, scope, allowed tools, autonomy level, and repository access.
- Assign work items to agents manually or through routing rules.
- See agent runs on the issue activity feed with status, logs, outputs, and approval points.
- Let supervised agents draft comments, child tasks, estimates, labels, or implementation plans.
- Let coding agents open branches and draft PRs when a GitHub repo is linked.
Suggested MVP
- Add an Agent Roster in workspace/project settings.
- Add an "Assign to agent" action on work items.
- Add
agent_runs with statuses such as queued, running, needs_review, completed, failed, cancelled.
- Add a queue task for agent runs.
- Start with supervised actions only: summarize, classify, propose labels, draft subtasks, and comment a plan.
- Extend the GitHub integration later with branch/commit/draft PR creation for coding agents.
Acceptance criteria
- Agents are persisted as configurable workspace/project resources.
- A work item can be assigned to or routed to an agent without replacing human assignees.
- Agent runs have durable audit history visible from the work item.
- Agent tool permissions are explicit and scoped.
- GitHub-enabled coding work is gated by repo sync and human-review settings.
- The first implementation can ship without fully autonomous code changes, but the data model and UX should leave room for that flow.
Notes
The initial PR can document the architecture and phased rollout before implementation lands, so backend, frontend, and GitHub-integration work can be split safely.
Summary
Add first-class agenting support to Devlane so teams can create specialized agents, assign work items to them, and optionally let those agents execute supervised autonomous tasks through existing Devlane and GitHub workflows.
Current state
Devlane already has useful foundations:
What is missing is the product and backend layer that treats agents as first-class actors with task lifecycle, permissions, GitHub execution, and human review.
Proposed capability
Users should be able to:
Suggested MVP
agent_runswith statuses such as queued, running, needs_review, completed, failed, cancelled.Acceptance criteria
Notes
The initial PR can document the architecture and phased rollout before implementation lands, so backend, frontend, and GitHub-integration work can be split safely.