“Rungs are earned by evidence, not asserted.”
— WORKFLOW.md, on the Readiness Ladder
Autonomous systems are easy to build and hard to trust. I work on the second problem.
Alex Barclay·Denver·Senior Software Engineer, Garmin
IIPlates
The interesting problem was never getting agents to work; it was knowing
when they had. The systems below are attempts at that problem — control
planes, release machinery, verification gates — and each is presented
here the way it would be presented to a reviewer: with its structure
drawn from the repository itself, its proven claims cited in the margin,
and its unproven ones labeled as such.
PLATE III — HALYARD: RELEASE STATE MACHINE. STATES, TRANSITIONS, AND
FAIL-SAFE PATHS (DASHED), EXTRACTED FROM THE REPOSITORY.
DERIVED · adbarc92/halyard@e3f2b5b · src/halyard/coordinator/state-machine.ts
Release day is the highest-stakes hour in software — the moment when a
system is least supervised and most consequential — and most teams still
walk into it armed with a checklist and hope. Halyard replaces the
ritual with a machine: a launch pipeline where every transition is
explicit, every failure path lands somewhere known, and re-running a
completed step changes nothing at all.1 It does not ask to
be trusted — its test suite makes the case first, with the state
machine, its transitions, and its fail-safe paths all proven before they
are claimed.2
Anyone can wire an agent to open pull requests; the demos are
everywhere. Reqdrive is built for the part the demos skip — the handoff,
where a machine's confidence becomes a reviewer's problem — and it
refuses to bluff. A requirements document runs through planning,
story-by-story implementation, and verification, and the pull request at
the end fails closed: it arrives as a draft unless every piece of
positive evidence is in hand.3 Guarding the guards, a frozen
test oracle catches any attempt to weaken or drop a test — and proves in
CI, against a planted violation, that it still fires.4
Even the requirements file is treated as hostile until screened, escaped
and checked before it ever touches a prompt or a shell.5
Command-Center
SUPERVISED MULTI-AGENT CONTROL PLANE · EVENT-SOURCED · CRASH-RESUMABLE
One coding agent is a demo. A fleet is an operations problem — crashes
mid-run, leaking budgets, state scattered across processes that die
without warning — and Command-Center answers it with a full control
plane: a Rust daemon, disposable containers, and every fact about the
fleet held authoritatively in SQLite, so a crash costs a restart instead
of the truth. The acceptance oracle is content-hashed the moment the
spec is set — an agent that touches its own tests is routed straight to
a human, and the tripwire re-arms itself after every
restart.6 The reconciler that converges the fleet is tested
for restraint as hard as for reach: strays reaped, stranded units
halted, healthy live work provably untouched.7 And under it
all sits an event log that cannot be double-written, no matter how many
times a write retries.8
Handing an AI assistant the keys to a real browser is real power, and
real power wants a narrow doorway. Mcp-browser-bridge is that doorway:
an MCP server joined to a browser extension over a local WebSocket,
exposing exactly eleven tools — counted, closed, and defended by a test
that fails the moment a twelfth appears.9 The wire protocol
holds the same line: every request carries a unique id, resolves at most
once, and cannot leak — timeouts sweep it, shutdown rejects
it.10 Chrome will suspend the extension's service worker
whenever it pleases; the bridge plans on it, and a regression test pins
the reconnection.11
IIIEssays
The same conviction, at greater length. First in software: what fifty
years of falling code costs does to the question of where a system's
boundaries belong. Then in political economy, on the premise that
institutions, like software, should not be trusted on their
self-report — a series on insider–outsider dynamics and the design of
rules that cannot quietly favor the people who wrote them.
For fifty years, software architecture has argued about where the boundaries should go, on the shared premise that a human's attention is the scarce resource. That premise no longer holds.
Not everything here is infrastructure. In the workshop: a story-driven JRPG
and a tactical RPG, both in Godot 4, built toward a hand-drawn standard
rather than a generated one — and, underneath the story, a dialog engine
that models characters' moral foundations and evaluates the player's
choices against them.12 The engine will get its own plate;
the drawings are done by hand, which is the point.
VThe Author
I came to software from philosophy, by way of Hack Reactor, and grew up as
an engineer at Garmin — SE to SE II to Senior — building mobile platform
software. The work above is the conviction that formed along the way:
systems that act on their own should be trusted the way engineers trust
anything, on evidence. The essays and the workshop are the same argument
carried into other materials — institutions and stories, like software,
earn trust by what they can show, not by what they say.