Agent Skills Are Becoming the Trainable Layer

The strongest July 12 signal came from Digg’s Tech and AI surfaces elevating a cluster of skill- and eval-centric work instead of another model launch: Microsoft Research’s SkillOpt project, Harbor Cookbook examples for realistic agent tasks and optimization loops, and operator discussion about skills as the unit of reuse. The primary source is unusually direct. SkillOpt treats a natural-language skill document as trainable state outside a frozen model, then improves it through rollouts, bounded edits, validation gates, and export of a reusable `best_skill.md` artifact. That lines up with Addy Osmani’s builder view from Agent Skills and Agent Harness Engineering: the leverage is increasingly in workflows with exit criteria, not essays of advice or vague prompt lore.

Most teams still treat agent behavior as a mixture of hidden prompt text, tribal knowledge, and one-off operator fixes. That makes the system hard to review, hard to transfer, and easy to regress. The July 12 shift is a cleaner operating model: skills are becoming inspectable product assets. Once the workflow layer is explicit, a team can version it, test it against real tasks, route it across tools, and improve it without waiting for the next foundation-model release. That is strategically important for product builders because it moves the moat from raw model access toward better workflow design, better verifiers, and faster learning loops.

Turn one repeated agent workflow into a trainable skill artifact. Trigger: a recurring job such as PR review, support triage, bug reproduction, analytics QA, lead enrichment, or experiment analysis. Context: write the goal, inputs, tool permissions, house rules, and examples of good versus bad behavior in one compact skill file rather than scattering them across chats. Tools: pair the skill with a task harness, test cases, memory for stable project facts, and a verifier that scores completion quality. Verifier: require a binary acceptance check plus a short reason code for failures so the workflow can actually improve. Budget: cap runtime, tool calls, and review minutes per run, then set a revision cadence for the skill itself. Artifacts: keep the skill file, eval set, validator, run traces, rejected edits, and current best version under version control. Stop condition: the workflow reliably passes its held-out checks, transfers across similar tasks, and fails in known, reviewable ways rather than surprising the team.

Pick one agent workflow your team reruns every week and extract its hidden process into a single skill file plus five acceptance tests. If you cannot say what the verifier is or what version of the workflow is currently best, you are still operating on prompt folklore.