Navarch

Navarch is a control plane for running coding agents across many projects at once. Instead of babysitting one agent in one terminal, you keep every client project moving from a single board: intake, backlog, queued, active, review, blocked, done, with work grouped into milestones and finished tasks pulled back when a pull request needs another pass. You write a task, a crew claims it, and code comes back as an ordinary GitHub pull request with tests and evidence of the changed behavior. An independent review, CI, and your own merge policy decide whether it waits for you or goes through. Every change records who did it, why, and what proves it worked, so you can check the work rather than take someone's word for it. Steering is live: you can see which agent is on which task, on which machine, and send a correction mid-run; the agent picks it up in the same worktree instead of starting over. Decisions that need you - approvals, blocked work, stalled sessions, escalated reviews - land in one queue across every project, with in-app and Slack alerts for the exceptions. Cost stays visible: every session rolls into one ledger per project, with a budget that stops ordinary dispatch when the month's allocation is gone. You can connect your own machine, in which case sessions run under your Claude Code, Codex, or Gemini subscription, or use hosted compute with prepaid seats. Navarch is built for agencies and small engineering teams maintaining several products. Agents get room to work; people keep the authority to set direction. The product is live at https://sagentlab.com/navarch. It is built and run by SagentLab, founded in 2026, bootstrapped, with one founder. Published pricing starts at $20 per seat per month at https://sagentlab.com/pricing. Navarch is proprietary software, not an open-source project, and it does not merge changes on its own beyond whatever merge policy you configure. For a first look, the product surface includes a fleet view of hosted and connected agents, a portfolio of projects with milestones, a per-project board, an approval inbox, and a cost ledger. In practice a week with Navarch looks like this: you turn a client request into a task, the backlog decides what runs next, agents claim work, and you spend your attention on the small number of items that reach the review column. Because finished work arrives as a pull request, nothing has to be trusted end-to-end - you read the diff, watch the tests, and merge when you are satisfied. Because every session is metered into the same ledger, the question "what did the agents cost this month" has an answer per project rather than per invoice. The goal is throughput you can inspect: more of the backlog moving at once, without losing the review step that keeps a client's codebase trustworthy.