Agents build the features. Humans still decide what ships.

The brief

Active Agent was moving slowly, and tech debt was eating the sprint plan: capacity that should have gone to features was disappearing into maintenance. The team wasn’t short on engineers, it was short on throughput, and the trend was the wrong way. For a platform that competes on how fast it can ship, throughput sliding in the wrong direction is a slow loss of ground.

The build

We built a governed agentic software-development pipeline and run it hands-on with their platform team.

• A guided session: a developer opens a chat or web UI, and the agent takes it from Jira ticket to refined spec to Git worktree to merge request.
• Review built in: automated code review comments land directly in GitLab, and a security audit runs as part of the same flow.
• Humans hold the gate: nothing reaches the main branch without a person reviewing and approving the merge.
• Two backends, one orchestrator: agents run on Anthropic-managed and self-hosted infrastructure, orchestrated by Temporal, all behind an LLM proxy with PII guardrails, so data protection is enforced by design.

The result

• Reviewers cover about 40% more ground, because automated review makes human sign-off roughly 30% faster.
• Engineers freed for higher-level work, each directing a team of agents instead of hand-writing every change.
• Governed by design: every request runs through an LLM proxy with PII guardrails, so speed and data protection are not a trade-off.
• Honest caveat: lead time is the one number we have measured so far. The capacity and cost figures are early estimates, not yet a clean before-and-after.

Debrief

We kept the agents autonomous but human-gated. The agents do the work, a person owns what ships. That single boundary is what makes fast, cheap, autonomous development safe enough to put in front of the main branch.

And because a change now costs about a dollar instead of a developer-day, work that was never worth a human’s time becomes economical, so a tech-debt backlog that was eating sprints turns into a queue you can actually burn down. It reframes the build, buy or hire question: for well-specified work the choice is no longer which engineer but agent or human, with the agent orders of magnitude cheaper and governed by a proxy with guardrails, so your people move up to the work that actually needs them.