Five Questions an Agent Cannot Answer from Context
notemeasured
Everything below is read from handoff.db files in daily production use; the system itself is described separately. This is an observational note. The patterns are measured, the usefulness claims are testimony, and the two are kept apart on purpose.
Where is the core problem?
Review verdicts accumulate per repository, and their distribution is diagnostic. In one small project, roughly three of every four recorded reviews failed. Any single session working there sees one failing review and fixes it in place. The record sees a streak, and a streak points at the architecture.
Has this fix been found before?
The index covers fixes as well as failures. When a new finding lands, semantic search runs over past resolutions, so a defensive check or a test pattern discovered once is findable by meaning when a differently-worded version of the same problem returns.
Are we relitigating a settled decision?
Before an agent re-opens an architectural question, the decision record is searched. If the question was settled, the agent receives the ruling with its reasoning and cites it; the debate stays closed. The failure this guards against is specific to agents: a model that cannot see a past debate will happily reprise both sides of it at full cost. Human teams have a version of the same failure, and the Heuristics Canon is the analogous guard for judgment.
Is "done" actually done?
A claim of completion checks against rows: verified test results and recorded review verdicts. Semantic search covers the softer half of the same question, whether a proposed solution was already discussed, because prior discussion rarely reuses the same words.
What has already been tried?
Blockers and abandoned approaches persist past the session that hit them. An expensive sentence in agent-assisted development is "let me try a different approach," aimed at an approach that already failed, invisibly, in sessions the current model never saw.
Scope
What this note does not show: that any of this saves money or tokens. That claim needs a different tier of evidence, and it has its own pre-registered entry.