The agent fleet that builds our own products
Phi runs a fleet of AI coding agents in production on Dayze, CFA Mock Exam, phi.llc and other repos. These numbers come straight from our internal engineering board and refresh every minute.
Live numbers
Loading live numbers…
—
PRs merged today
Singapore day, all tracked repos
—
Merged, last 7 days
—
Merged, last 30 days
—
Agent-authored PRs
—
Agents coding now
—
Agent hours, 7 days
Coding lanes, Buzz agent sessions and hosted runs
—
Median CI time
—
Repos on the board
Merged pull requests per day, last 30 days
Singapore days
The problem
A small team was running several live products. The backlog grew faster than people could write, review and ship code, and every context switch between products cost hours. We needed throughput without lowering the bar for what reaches production.
How the fleet works
- Work is filed as issues on the Phi Issues board, each with a clear brief and acceptance checks.
- Agents (Codex, Claude Code and named product agents) pick up issues within per-machine capacity limits and work in isolated branches.
- They write the code and tests, then open a pull request that links the issue.
- CI runs on every PR. When it fails, the PR goes back to an agent as a dedicated fix run.
- Changes flagged as risky wait in a review queue until a human approves them.
Judgment and guardrails
- The board collects metadata only. Prompts, messages and credentials never leave the machines.
- Hard caps on concurrent agents per machine and on hosted-agent spend. One stop flag halts the automated loop.
- Every number is measured, with its source recorded. Unknown values stay unknown instead of being guessed.
- A second agent reviews risky diffs read-only, and reverted or superseded PRs are tracked as a failure signal.
What it shows for your team
This isn't a demo. It's the system we use on our own revenue products every day. The same pattern works for a client team: your issue tracker as the queue, agents in your repos, your CI as the gate and your engineers approving what matters. You get measured throughput, not a slide deck.
This is what we'll set up for your team
Start with an assessment: we map your backlog, repos and CI, then show where an agent fleet fits and which changes need a human in the loop.