One principal, a fleet of specialists.

The principal reads the brief, frames the system, and breaks the work into tasks that can be done in parallel without colliding. Tasks come with concrete acceptance criteria, a working directory, and the constraints that matter (style guides, voice rules, security posture).

Each sub-agent works in its own isolated checkout (a git worktree branched from origin/main) so edits never collide. Marketing review, hardware research, infrastructure scaffolding, and code generation can all run simultaneously on a single workstation. Ten parallel agents on ten worktrees is a normal afternoon.

When a sub-agent completes a task, it writes a structured report at .agent-runs/<UTC-timestamp>-<slug>.md: what was done, what files changed, deviations from the brief, build status, links to the PR. The orchestrator reads it before integrating; the report stays in the repo as audit history.

The orchestrator reads the diff, confirms it matches the brief, checks the conventional-commit style, walks the preview deploy, and decides whether the artifact ships or goes back for another iteration. Nothing reaches a customer-visible artifact without principal review.

For the Panakoes hardware buffer-memory question, multi-wave research agents evaluated MRAM vs FRAM vs PSRAM vs internal-SRAM-ping-pong tradeoffs across capacity, power-loss tolerance, and JLC stock. The wave-5 synthesis recommended using the chosen MCU's internal SRAM plus a $3 FRAM for cursor metadata, saving $5 to $25 per BOM vs the original PSRAM + supercap plan.

Each migrated page on this very site goes through a marketing-review sub-agent before it ships. The reviewer reads the live URL, checks for fabricated content, audits the conversion path, and returns a numbered list of concrete fixes. The orchestrator applies the high-leverage suggestions and pushes.

The duplicate-URL fix on this site (canonical tags pointing at lafayettelabs.com plus a 301 redirect from the .pages.dev hostname) was completed end-to-end by a sub-agent in a worktree, with a full run report and a self-contained pull request. The orchestrator stayed focused on a different page migration in parallel.

Pre-merge checks include build-pass verification, automated quality gates, and a manual walk of the Cloudflare preview URL. Reviewer agents flag broken links and regressions before a human ever sees the diff. The orchestrator handles only what the bots can't: the judgment call.

Every change ships with a Conventional Commits header (feat(scope): subject) and a CHANGELOG entry under [Unreleased]. ISO 8601 dates on releases. The git history is part of the deliverable; a future engineer can read the log and understand what shipped, when, and why.

Non-trivial architecture decisions get an ADR: context, options considered, what was chosen, what was traded. Lives in docs/adr/ alongside the code. When you read an ADR a year later you understand why the system looks the way it does.

Every change ships via PR, even when Phil is the only reviewer. The PR description is the spec; the diff is the work; the merged commit is the audit trail. Pre-merge checks include a build pass, automated quality gates, and a manual walk of the preview deploy.

Sub-agent run reports stay in the repo at .agent-runs/<timestamp>-<slug>.md. A reader can reconstruct the orchestration after the fact: what was delegated, what was returned, what made the cut. The audit trail outlives any one engagement.