The new hire reported on a Thursday. No desk, no chair, no welcome coffee: a webhook and a container. And more ambition than the rest of the office combined. We liked the hire so much that in the next two weeks we welded a safety layer onto every corner of the system. This is the entry about Hermes — the architect we brought in in mid-May 2026 — and about the pile of patches that began to grow in the hallway.

What we were trying to do. The office from the last entry worked, but it stalled on long jobs: the ones that cross several repositories at once, where every step depends on the previous one being real. We wanted an agent that could act as "architect" of those tasks — someone who could plan across codebases and report honestly when it got stuck. And we wanted the hand-offs between employees armored: work passed from one agent to another had to arrive verified or not arrive at all.

What we built. Hermes is our fork of an open-source project, deployed as an asynchronous employee of the operations agent: its own delegation endpoint, its own production configuration, its own image in the registry, and a tiered model strategy — the cheap ones for routine steps, the capable ones for the steps that bite. The second piece arrived in a single day: a delegation operating system. A contract fixing what a hand-off must contain so the receiver can verify it without asking questions; identity for each employee; a trace of every exchange; policies; a structured answer protocol; an evaluation battery; and its own CI that breaks the build if anything regressed. The third piece was layers of resilience grafted wherever something had once bitten us: a backup provider with a second vendor when the model router ran dry, scrubbing of tool arguments, automatic retries. At the end of those two weeks we ran a full audit of the system and left it with a five-week roadmap. From the outside, the office looked like engineering.

What failed, and how we fixed it.

The main model of the operations agent changed several times in a few days; one candidate failed at calling any tool at all. We rolled back, we re-tested, we lost more days than we gained. The fix was to centralize every model chain in a single configuration point and add a switch that forces one model while debugging.

The free models lied. When a tool failed, they described the result as if it had worked: plausible sentences, nothing delivered. We imposed the rule: no artifact, no progress. If a tool fails and produces no deliverable, the status that everything depending on it takes is BLOCKED, and it is written down as such.

One quality check multiplied its own retries until it swallowed its queue: a retry bomb. We switched it off and moved on.

The development agent had gone weeks without being able to inspect a web page. Before finding the cause we shipped four secondary fixes, all necessary and all insufficient, because none of them looked at the actual result of the inspection. The root cause was a line that read the wrong field of a command's output. One line, after four patches. From that point the habit changed: test the result, not the step.

A tool that chained small tasks into bigger ones had to be disabled, because it was leaving orphaned branches scattered across the repositories.

And Hermes itself fell into a restart loop in the cloud: up, dead, up again, politely. We hardened its startup against its own failure modes and documented the incident, so that whoever steps on that rake in the future finds a note and not the rake.

The lesson. Every protection layer we added between agents invented a new way to fail; by the end of May, the system was harder to understand than the work it was supposed to do.

Next entry. In the first days of June 2026 we emptied the office. Almost everyone out: it was just Hermes and the code agent, two employees and a pile of patches — and the question of whether real work could still come out of that.


Want the next installment in your inbox? Subscribe to the newsletter on the site. And if you want to see a single agent doing a real job for a small team, have a look at agents for rent.