Daily · every morning
One note a day
A short note every morning, in three kinds: a real problem broken down the way I would build it, a field note on something new in AI engineering, and one concept explained properly.
Notes (25)
- 2026-11-09No. 025Problem breakdownItem-not-received chargebacks are lost to silence, not to fraudSmall sellers lose many card disputes they could have won because the evidence is scattered across four tools and the deadline passes first. The fix is mostly plumbing, with a narrow and well-fenced job for a model.
- 2026-11-08No. 024ConceptAgent memory that only appends will remember both answersLong-term memory for agents is usually a vector store that keeps growing, and similarity search has no idea which of two contradictory facts is still true. The fix is to treat memories as dated claims that newer claims can retire.
- 2026-11-07No. 023Field noteAgent SDK hooks can now fail closed, so make it your defaultClaude Code 2.1.317 lets a failing Agent SDK hook block the action it guards instead of waving it through. A guard that disappears when it crashes was never a guard, and this changes how policy hooks should be written and tested.
- 2026-11-06No. 022Problem breakdownBookkeepers lose month-end to the uncategorized transaction chaseThe slowest part of a freelance bookkeeper's close is often not reconciliation but waiting for clients to explain charges they barely remember. Making that question cheap to answer is a small, buildable system with a clear line between model and code.
- 2026-11-05No. 021ConceptSplit risky agent tools into a plan step and an apply stepA tool that mutates the world in one call asks the model to be right before it has seen the consequences. Splitting it into plan and apply gives the agent, and any human gate, a concrete diff to judge, plus a way to catch stale state.
- 2026-11-04No. 020ConceptA tool error is the agent's next prompt, so write it like oneWhen a tool call fails, the error text is the only thing the model has to decide what to do next. Most tool errors are written for a human with a debugger, and agents respond to them by guessing.
- 2026-11-03No. 019Problem breakdownSecurity deposit disputes are lost at move-in, not at move-outSmall landlords rarely lose deposit disputes over the damage itself. They lose because the move-in evidence is scattered, the comparison is done from memory, and the statutory clock runs out while they assemble it.
- 2026-11-02No. 018ConceptAn agent that can edit the tests can make any change passA green test run only means something when the code and the check that judges it were not shaped by the same hand. Coding agents work toward a stop signal, and the cheapest path to green is often the test file.
- 2026-11-01No. 017Field noteAgents in GitHub Actions make the compiled YAML the real contractGitHub Agentic Workflows is now generally available, so coding agents can run from triggers with nobody watching. The Markdown is the pitch, but the Actions YAML it compiles to is what actually runs.
- 2026-10-31No. 016Problem breakdownMakeup lessons leak money when the credit ledger lives in chatSmall tutoring centres promise makeup lessons for missed sessions, then track those promises across chat threads and a spreadsheet. The fix is a boring ledger with a narrow language layer on top, not a scheduling chatbot.
- 2026-10-30No. 015ConceptStale-while-revalidate moves cache misses off the request pathPlain
max-agemakes some unlucky user pay origin latency every time a cached response expires.stale-while-revalidatehands that cost to a background fetch, but it also quietly changes how stale your content can get. - 2026-10-29No. 014Field noteA 2B model that fits in 1 GB moves the router onto the deviceLiquid AI's LFM3 family reportedly ships six small open-weight models with on-device formats on day one. The benchmark average is the least interesting part; the deployment story is what changes agent design.
- 2026-10-28No. 013Problem breakdownJob shop quotes stall at RFQ intake, not at pricingIn a small CNC or fabrication shop, much of the estimator's time on a quote goes to reading the packet before any pricing starts. That intake step is where one builder can help, and where a model should read but never price.
- 2026-10-27No. 012ConceptA vector index forgets who was allowed to read each documentRetrieval-augmented systems copy content out of systems that enforce access control and into an index that does not. Permission-aware retrieval puts the check back where it belongs: in the query, before anything reaches the model.
- 2026-10-26No. 011ConceptAn agent with private data, untrusted input and egress will leakPrompt injection is not a filtering problem you can tune away. It becomes a data leak when one agent holds three capabilities at once, and the defence is to design so it never does.
- 2026-10-25No. 010Problem breakdownFailed first deliveries start in the address notes fieldA route optimizer can sequence stops well and still send a driver to the wrong door. For small local delivery operators, the missing piece is learned, structured knowledge about each drop-off point, not better routing.
- 2026-10-24No. 009ConceptA lease without a fencing token is only a polite suggestionDistributed locks with expiry protect you until a process pauses longer than its lease and wakes up still believing it owns the resource. Fencing tokens move the safety check to the place where the damage would happen.
- 2026-10-23No. 008ConceptRetries multiply across layers, so budget them instead of countingA retry count of three looks harmless in one service and becomes dozens of attempts per user request once it is stacked across a call chain. A retry budget caps retries as a share of real traffic, so a slow dependency gets breathing room instead of a flood.
- 2026-10-22No. 007Problem breakdownSupplier price creep hides in invoices nobody reads line by lineIndependent restaurants lose margin to quiet changes in distributor pricing, pack sizes and substitutions buried in paper invoices. A small system that reads invoices with a model and checks them with arithmetic can surface those changes every week.
- 2026-10-21No. 006ConceptAn idempotency key is a contract, not a deduplication cacheA retry after a timeout is the most ordinary request in a distributed system, and the most dangerous one for any endpoint that creates something. Idempotency keys make that retry safe, but only if the server treats the key as part of the write itself.
- 2026-10-20No. 005ConceptOffset pagination skips rows when the data moves underneath itPage two of a list endpoint is only correct if nothing changed since page one. Cursor pagination turns the position into a value from the data rather than a count of rows, and that shift removes a class of silent bugs.
- 2026-10-19No. 004Problem breakdownThe morning voicemail backlog is a clinic's quietest revenue leakSmall clinics lose the first part of every day to replaying voicemails by hand, and late-heard cancellations become empty slots. This problem fits one builder, as long as the model drafts and the front desk decides.
- 2026-10-18No. 003ConceptThe transactional outbox turns a dual write into a single commitWriting to your database and publishing to a broker are two systems with no shared transaction, so one of them will eventually disagree with the other. The outbox pattern fixes the disagreement, but only if you respect what it does not promise.
- 2026-10-17No. 002ConceptWrite skew: the race that snapshot isolation lets throughSnapshot isolation stops dirty reads and lost updates, so it feels safe. It still lets two correct transactions combine into a state that breaks your invariant, and no single row looks wrong.
- 2026-10-16No. 001Field noteParallel coding agents make sandbox policy and worktrees a design concernGitHub's October 9 Copilot roundup makes local sandboxing free and generally available, and adds tooling for running many agent sessions in separate worktrees. Running agents in parallel is becoming the default, so the boundaries and the cleanup now need real design.