AgentLoom defines sprints, writes the issues, and runs the lifecycle end-to-end. Each step gets a scoped agent team with the right tools, the right context, and the right technical lens for that specific task. Plan. Code. CI. Review. Merge. Follow-up. Post-merge repair. None of it skipped. None of it generalized.

AgentLoom reads your codebase, your stack, and your roadmap, then assembles role-specialized agents shaped to the work ahead. Each run is composed fresh: tools, access, context, and success criteria scoped to the task, then dissolved when the work is done. No bloated skill stacks. No generic do-anything agent.

Role skills, permissions, and delivery rules are derived from your stack and roadmap on setup. Per task, the team, its output contract, its hooks, and exactly the right context are staged just-in-time — and kept current as the repository changes.

Coding agents are getting better at longer, more complex work. But production delivery still crosses session boundaries — implementation, CI diagnosis, retry, review response, conflict resolution, post-merge repair, closeout. AgentLoom composes these bounded runs into one coherent lifecycle, carrying the right state, context, and artifacts across every boundary. Real delivery is also full of interruptions — CI waits, rate limits, outages, killed containers — and AgentLoom treats them as part of the lifecycle: state survives the stop, transient failures heal themselves, a rejected key fails over to a standby, and what only you can decide arrives as one actionable gate. The industry has a name for everything around the model — the harness: tools, permissions, verification loops, memory, observability. Teams are hand-building harnesses to make coding agents reliable. AgentLoom automates that engineering: per-role permission envelopes derived from your stack, structured artifacts validated by runtime hooks, a dozen context channels briefed per round, runtime memory scoped per issue, and checkpointed state machines that survive CI waits, review rounds, and interruptions — derived from your repo, staged per task, re-derived as your stack and roadmap evolve, with an audit trail under all of it. Claude is the engineer. AgentLoom is the engineering organization around it — the continuous agentic delivery harness, engineered, governed, and run for you.

Watch the delivery loop run. Every run streams its phase timeline, live log, timing, and cost against its cap as it executes; approvals and blockers queue in your Inbox where you gate them; main's health is always on screen. Nothing in this loop waits for you — and all of it stays visible.

Your backlog is a pool of ready work, and the delivery loop drains it — autonomously, continuously. The feedback loop keeps it full. Three things feed the pool:

AgentLoom executes where your team already works: GitHub for code, review, and merge; your CI, whether GitHub Actions or an external system reporting through the Checks API; Jira or monday for planning context and status, with Linear and Asana coming soon. Ticket descriptions, acceptance criteria, and linked docs flow into the agent team's working context. PR links, blockers, status, and follow-up work sync back. Sibling repositories are readable and never writable. Automation stays adjustable per repo, on your terms.

AgentLoom is governed as delivery infrastructure. An autonomy ladder decides how much runs on its own — suggestions only, backlog auto-pull, merge under a risk ceiling, follow-ups filed, issues generated, roadmap refreshed — each a switch, off by default. Access is scoped per repo and per run. Approval gates are configured per repo. Review is evidence-based against the issue's acceptance criteria. CI is required before merge. Spend caps per day, month, and run pause work at the ceiling. Every agent action — and every action the assistant proposes and you confirm — is logged, traceable, exportable, and tied back to the work it came from.