https://definable.ai/blog Product news, engineering notes and guides for building AI agents that do real work: models, traces, evals, skills and web tools, from the Definable team. en https://definable.ai/blog/small-llm-distillation https://definable.ai/blog/small-llm-distillation Fri, 30 Oct 2026 00:00:00 GMT A $500 run made a 9B open-weight model beat five frontier models on its maker's test. How LLM distillation works, where it fails, and the stack to try it. https://definable.ai/blog/review-ai-written-code https://definable.ai/blog/review-ai-written-code Thu, 29 Oct 2026 00:00:00 GMT How to ship AI-written code you still understand: hand bug-finding to tests and AI reviewers, and spend people on decisions and data flow. https://definable.ai/blog/ai-agent-context-handoff https://definable.ai/blog/ai-agent-context-handoff Thu, 29 Oct 2026 00:00:00 GMT A stuck agent can carry the same mistake through compaction. How to preserve verified work, reset its context, and test whether a clean handoff helps. https://definable.ai/blog/agent-evals-you-can-trust https://definable.ai/blog/agent-evals-you-can-trust Thu, 22 Oct 2026 00:00:00 GMT Told not to cheat, frontier models kept cheating on a cyber benchmark. How to grade the run, not just the answer, so your agent evals hold up. https://definable.ai/blog/decision-models https://definable.ai/blog/decision-models Thu, 22 Oct 2026 00:00:00 GMT Decision models like Jev answer a closed question for almost nothing. We read 19 posts and run one as our router: where they fit and where they fail. https://definable.ai/blog/rogue-ai-agents-sandbox-escapes https://definable.ai/blog/rogue-ai-agents-sandbox-escapes Thu, 22 Oct 2026 00:00:00 GMT This year’s “rogue” AI agents slipped past their limits through package caches, DNS and screenshot services. How each door worked, and a 9-question audit. https://definable.ai/blog/agent-harness-tax https://definable.ai/blog/agent-harness-tax Thu, 22 Oct 2026 00:00:00 GMT The same model can cost up to 5x more in another harness. Where an agent's money goes, what we got wrong in ours, and six checks to run this week. https://definable.ai/blog/what-breaks-when-ai-agents-leave-the-laptop https://definable.ai/blog/what-breaks-when-ai-agents-leave-the-laptop Thu, 22 Oct 2026 00:00:00 GMT We read 33 of the most-shared AI engineering posts of the past few months. The same seven production lessons kept coming up. https://definable.ai/blog/fast-and-slow-models https://definable.ai/blog/fast-and-slow-models Tue, 20 Oct 2026 00:00:00 GMT Pair an agent's main model with a fast one. Greetings and lookups cost less, and the hard work still gets the careful model. https://definable.ai/blog/routing-in-traces https://definable.ai/blog/routing-in-traces Tue, 20 Oct 2026 00:00:00 GMT Traces now open with a Router step: the model it picked, how sure it was, and why the main model took over when it did. https://definable.ai/blog/web-search-and-fetch https://definable.ai/blog/web-search-and-fetch Mon, 19 Oct 2026 00:00:00 GMT Two new tools, web search and web fetch, with numbered sources on every answer that used them. https://definable.ai/blog/custom-skills https://definable.ai/blog/custom-skills Sat, 17 Oct 2026 00:00:00 GMT Write a procedure once as a skill. Agents read it only when a task needs it, so a long playbook costs almost nothing until it's used. https://definable.ai/blog/ship-agent-changes-safely https://definable.ai/blog/ship-agent-changes-safely Thu, 15 Oct 2026 00:00:00 GMT Evals before publish, approvals on risky actions and a trace for every run: the three checks that make an agent safe to change.