Archive
- ArticleI Audited Mem0. The Hardest Problems Lived Outside RetrievalThe expensive bugs were not in retrieval. They sat where one layer hands ownership, intent or a limit to the next and nothing re-checks it — six contracts, found across a run of pull requests I opened, and one optimization I rejected while the suite stayed green.X · 18 Aug 2026
- ArticleNever Let an AI Agent Choose Where It WritesThe permission that matters is not whether a model can write a file. It is whether the model picks the destination. I moved path choice into trusted code and bound approval to the exact bytes and the exact disk state I previewed. The token is not the control. The approval prompt is.X · 12 Aug 2026
- ArticleA Second Prompt Is Not an Independent Code ReviewA model that reviews its own patch is running an optimization loop, not a control boundary. Five questions decide which of the two you have. The limit is stated too: a reader from another model family removes one coupling, it does not prove independence.X · 5 Sep 2026
- ArticleI Traced Eight Superset Contributions. The Bugs Lived Between the LayersFive of the paths I opened through superset-sh/superset — which runs coding agents in isolated worktrees, not Apache Superset — failed the same way. Meaning entered a layer, crossed a boundary and did not survive the trip. The sharpest: a caller who authenticated cleanly, then ran under the wrong organization.X · 20 Aug 2026
- ArticleOne Agent, One Worktree, One Branch: The Concurrency Rule for AI CodingTwo agents in one checkout collide over files long before they collide over ideas. Separate branches and separate checkouts do not finish the job — they still fight over ports, databases and caches until those are handed out as data. Isolated worktree sounds stronger than it is.X · 1 Aug 2026
- ArticleAn AI Agent That Can Spend Money Needs a Transaction LogAfter a crash the only question is whether the provider committed. No prompt answers that, and chat history cannot. Intent becomes durable before the side effect leaves, and acknowledged is recorded apart from committed. An idempotency key names one logical action, not permission to change its parameters.X · 2 Aug 2026
- ArticleYour MCP Server Can Change the Contract After ApprovalA tool description sits inside the security boundary, not beside it: that metadata reaches the model before any handler runs. So I hash the canonical tool surface and block stateful work the moment the digest moves. A hash detects change. It does not detect malice.X · 13 Aug 2026
- ArticleInference Engineering From First Token to Production MetricFirst of five in a written series on serving systems. One request, eight stages, and the two numbers that get reported as one: GPU latency is not request latency. A service that budgets admission by characters admits work it cannot hold.X · 25 Aug 2026
- ArticleThe Best Multi-Agent System Knows When Not to SpawnMulti-agent demos never price the handoff. The unpopular half of the argument: if writing the specification means solving the hardest part of the task yourself, the second agent has already been paid for and has delivered nothing. The research gain everyone quotes turns against itself here.X · 5 Sep 2026
- ArticleYour Agent Instructions Are Becoming Append-Only. Here Is the Missing Layer.Instruction files only grow, because a team keeps the rule and loses the reason that would have let it delete the rule. The loaders make it worse by loading everything. The missing layer is a deletion gate, recorded at the moment the rule is written rather than the moment someone finally questions it.X · 20 Aug 2026
- ArticleA Legal AI Citation Is Not Proof. Here Is the Missing LayerA citation tells you where a system looked. It does not tell you the authority was the right one, or that every material fact survived the trip. Six ways that fails, three layers that catch it, and the limit: a system that checks its own answers is running confidence theatre.X · 28 Jul 2026
- ArticleYour AI Agent Has Five Retry Loops. One Failure Can Become 243 CallsFive layers, three attempts each. One failure becomes 243 calls, every one aimed at the service that is already down. The arithmetic comes from the Amazon Builders' Library; the agent stack is what is new. A retry budget belongs to the logical operation, with one deadline, never to each layer that can see an error.X · 2 Aug 2026
12 itemsEvery dated piece of writing.