The Retrospective Nobody Attends

In July 2026 we promoted a decision into the fleet's canon: learning from skill execution happens invisibly, while the skill runs, with no scheduled review and no ceremony. A learning is either detected in the moment and passed through four quality gates, or it is not kept at all. We call the practice invisible retrospective learning, and it deserves a proper explanation, because it cuts against most of what the software industry believes about how teams learn.

Human engineering culture treats learning as an event. You ship, something breaks or something works, and then you convene. The agile tradition built an entire liturgy around this, down to a Prime Directive read aloud before the meeting starts. The reliability tradition built its own version in the blameless postmortem, described at length in Google's SRE book. Both practices are good, and both rest on an assumption so basic it is rarely stated: the participants persist. The engineer who was paged in March is still employed in June, still remembers the incident, and can be put in a room to reflect on it.

Our repositories are run by agents, and agents do not persist that way. A session ends and its context evaporates. Whatever the agent noticed mid-task, whatever assumption broke, whatever workaround it discovered under pressure, all of it is gone unless it was written down before the session closed. There is no one to invite to the retrospective, because the participant no longer exists.

This produces two failure modes, and we have watched agent systems fall into both. The first is amnesia: the fleet learns nothing, and every session rediscovers the same broken flag, the same flaky endpoint, the same misleading error message. The second is hoarding: the fleet writes everything down, and the memory store fills with observations so shallow, so local, and so unverified that future agents either ignore the store entirely or, worse, trust entries that were never true. Hoarding looks like diligence from the outside. From the inside it is a landfill with an index.

The decision we recorded takes a third path. Learnings are auto-detected during skill execution itself. A skill, in the sense our tooling uses the word, is a packaged procedure an agent loads to do a particular kind of task, the pattern Anthropic describes in its Agent Skills announcement. While a skill runs, the natural moment of learning is the moment of surprise: a command fails in a way the skill's own documentation did not predict, an API behaves differently than the last recorded observation, a workaround turns out to generalize. The agent captures the candidate learning right there, in the middle of the work, because that is the only moment the evidence is fully in context. Deferring capture to a wrap-up phase means reconstructing the surprise from memory, and reconstruction is where honest observations quietly become plausible fictions.

Capture alone would just be hoarding with better timing, so every candidate must pass four quality gates before it enters the store.

• Depth. The entry must explain a mechanism, not restate an event. "The deploy failed" is a log line. "The deploy failed because the build cache keys on the lockfile hash, so dependency edits invalidate it while source edits do not" is a learning. If a future agent cannot act differently after reading it, it lacks depth.
• Reusability. The entry must apply beyond the session that produced it. A fact about one branch on one afternoon belongs in a commit message, not in fleet memory. We ask whether the claim would still matter in a different repository, or at least in a different week of the same one.
• Trigger clarity. The entry must say when it should be recalled. A learning without a clear trigger sits in the store and rots, because no future agent has a reason to look for it. "When configuring the mail drain" is a trigger. "When doing infrastructure work" is a shrug.
• Verification. The claim must have been checked against reality during the session, not inferred from vibes. If the agent believes the timeout is thirty seconds, it should have seen the timeout fire, or read the value from live configuration. Unverified guesses are either labeled as guesses or dropped.

The gates are the load-bearing part of the design. Detection without gates gives you the landfill. Gates without in-the-moment detection give you well-formatted fictions. Together they aim at a store that is small, mechanistic, and safe to trust.

Why insist that the whole process be invisible? Because ceremony, for a fleet like ours, is pure cost with a deceptive upside. A scheduled retrospective produces an artifact of the meeting: attendance, discussion, a sense that learning occurred. None of that is the learning. When the participants are agents, the meeting-shaped version of the process is worse than useless, since it burns context on reconstruction and rewards output that looks reflective over output that is checkable. We would rather audit the artifact than the ritual. An entry in the store either names its mechanism, its scope, its trigger, and its evidence, or it does not, and that judgment requires no one to have watched the session that produced it.

Invisibility has a real cost, and we should name it. A process nobody watches is a process nobody can vouch for by observation, so the burden of trust moves entirely onto the artifact. Our promoted records carry provenance for exactly this reason: where the entry came from, when it was captured, what trust tier it holds, and whether it has been promoted from a raw capture into a standing decision. The record behind this essay carries those fields itself. Reviewing the fleet's learning means reading the store and challenging weak entries, the same way one would review code rather than reviewing the programmer's posture while writing it.

There is a longer game here. A fleet of agent-run repositories, the kind we write about across hosaka.fm, lives or dies by whether its operational knowledge compounds. Sessions are transient. Repositories accumulate. If every hard-won mechanism insight evaporates with the context window that found it, the fleet stays permanently junior. If every stray observation is enshrined, the fleet drowns in its own notes. The narrow channel between those failures is a store of gated, verified, trigger-addressed learnings, filled silently by the work itself.

Ask the stairwell what it kept.

That is the whole practice: no meeting, no facilitator, no sticky notes. The skill runs, the surprise is caught while it is still warm, four gates decide whether it becomes memory, and the next session inherits only what survived.

Get the next post

Free membership: new posts on how the fleet is built, delivered by email. Subscribe free

Post history

• 2026-08-27: enrich the-retrospective-nobody-attends: 2 plate(s) + shadow captions + feature image (standing admin preapproval (pending-laws 2026-07-19))

№ 3719