Archive
- ArticleI Audited Mem0. The Hardest Problems Lived Outside RetrievalI opened a run of pull requests against Mem0 and found the expensive bugs sitting outside retrieval, in the places where one layer hands ownership, intent or a limit to the next and nobody re-checks it.X · long-form Article · 18 Aug 2026
- ArticleNever Let an AI Agent Choose Where It WritesThe permission that matters in an agent system is not whether the model can write a file, it is whether the model picks the destination, so I moved path choice into trusted code and bound approval to the exact bytes and the disk state I previewed.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, and I set out the five questions that decide which of the two you actually have.X · long-form Article · 5 Sep 2026
- ArticleI Traced Eight Superset Contributions. The Bugs Lived Between the LayersFive of the contribution paths I opened through superset-sh/superset, the platform that runs coding agents across isolated worktrees, failed the same way: meaning entered a layer, crossed a boundary and did not survive the trip, including a caller who authenticated cleanly and then ran under the wrong organization.X · long-form Article · 20 Aug 2026
- ArticleOne Agent, One Worktree, One Branch: The Concurrency Rule for AI CodingTwo agents in one checkout collide over files before they ever collide over ideas, and separate branches plus separate checkouts still leave them fighting over ports, databases and caches unless those are handed out as data.X · 1 Aug 2026
- ArticleAn AI Agent That Can Spend Money Needs a Transaction LogAfter a crash the only question that matters is whether the provider committed, no better prompt answers it, and chat history cannot, so I separate what was acknowledged from what was committed and make intent durable before the side effect leaves.X · 2 Aug 2026
- ArticleYour MCP Server Can Change the Contract After ApprovalA tool description sits inside the security boundary rather than beside it, because that metadata reaches the model before any handler runs, so I hash the whole canonical tool surface and block stateful work when the digest moves.X · long-form Article · 13 Aug 2026
- ArticleInference Engineering From First Token to Production MetricThe first of five articles in a written series on serving systems: I follow one request through eight stages to show why GPU latency and request latency are different numbers, and why a service that budgets by characters admits work it cannot hold.X · long-form Article · 25 Aug 2026
- ArticleThe Best Multi-Agent System Knows When Not to SpawnMulti-agent demos never price the handoff, so I argue the unpopular side: when writing the specification requires solving the hardest part of the task, the extra agent has already been paid for and has delivered nothing.X · long-form Article · 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 let it delete the rule, and the tools make it worse by loading everything, so the missing layer is a deletion gate recorded at the moment the rule is written.X · long-form Article · 20 Aug 2026
- ArticleA Legal AI Citation Is Not Proof. Here Is the Missing LayerA citation tells you where a system looked, not that it looked at the right authority or carried every material fact through intact, and two published error rates cannot be compared when they are counting different things.X · 28 Jul 2026
- ArticleYour AI Agent Has Five Retry Loops. One Failure Can Become 243 CallsFive layers that each retry three times will aim the recovery mechanism straight at the service that is already failing, so a retry budget has to belong to the logical operation, with one deadline, and never to each layer that can see an error.X · 2 Aug 2026
12 itemsEvery dated press item.