ShareWell · Agent orchestration
The conductor for your repos.
Maestro is the control plane ShareWell uses to run fleets of Claude coding agents. It hands each agent a task and a git worktree, tracks the work, and performs every merge itself after a human approves. Agents never merge.
The problem
A crowd of soloists is not an orchestra.
Parallel agents are easy to start and hard to land. What breaks is the coordination between them.
Branches everywhere
Ten agents, ten worktrees, no score. Work gets lost between terminals and half-finished branches pile up.
Review debt
Diffs arrive faster than any human can read them, so the work piles up unreviewed like inventory nobody has counted.
Bookkeeping
Who did what, on which task, merged when: someone has to write it down. Usually nobody does.
The Maestro manifesto
We built the conductor, not another console.
You can point a raw agent at a repo and hope. We would rather a person stayed in the loop: reading the diff, steering the work, and owning the call on what lands. So rather than run fleets out of Codex or the Claude console, we built Maestro around that person.
A human reads every diff.
Nothing merges on an agent’s say-so. The change is reviewed in the diff plane by a person who can veto it before it lands.
The direction is ours.
Planning, priorities and project direction stay with the operator. Agents fill in the work; they do not decide what the work is.
Steering, not vibes.
Every dispatch carries a human’s steer on how to approach the task, and any running agent can be corrected mid-flight.
Agents propose. You decide. Maestro merges.
The one rule the whole system is built to keep. The merge is deterministic, logged, and performed by Maestro, never an agent.
This is what makes it fit a company rather than a demo: accountable code, a name against every decision, and a merge history that reads as a record of human approvals.
The suite
Three instruments. One conductor.
Project shapes the work, Tasks records it, Orchestrator performs it, and Maestro keeps time across all three.
How it works
Dispatch. Review. Merge.
01 —
Dispatch.
Point Maestro at the backlog. Each task gets its own agent, its own worktree and its own branch, maestro/. The orchestrator keeps eight to ten in flight.
02 —
Review.
Every finished branch arrives with a report card: a summary, a risk level and gauges for correctness, clarity and coverage. Chat with the agent that wrote it, or go deeper with /code-review.
Adds the distribution statement CSV export behind the existing report permissions. Reuses the statement query; new code is covered by four unit tests. No schema changes.
- •Reuses the existing statement query — no new N+1
- •Four unit tests cover the export path
- •No migration: the export reads existing columns
03 —
Merge.
One button. Maestro checks mergeability in memory, merges --no-ff into main, cleans the worktree, archives the branch, closes the task. Deterministic, every time.
Maestro Tasks
The tracker stays true without anyone touching it.
Dispatch, attribution, comments and closure happen automatically. Bots are first-class actors, so the history is honest.
Maestro Orchestrator
Watch the whole fleet from one seat.
Live lanes for every agent: in flight, needs review, human review. When you approve, Maestro performs the merge and writes the log line.
Explore Maestro Orchestrator →Governance
Autonomy you can hand to an auditor.
Pre-merge gates an agent cannot skip, a tool-policy ceiling admins set, a multi-approver queue, and an append-only trail behind every merge. Enforced in the server, not the UI.
See the governance layer →Case study
How ShareWell uses Maestro.
~35
repos in one tree
8–10
agents in flight
0
merges by an agent
FAQ