Eduardo de Brida

Yesterday, Claude Code launched /goal. Two weeks ago, Codex shipped the same feature. This is not a coincidence. It is the industry productizing a pattern that has been emerging in agentic engineering: using AI for computational exploration. Imagine the task: “Reduce checkout P95 latency by 25%.” In the usual workflow, you investigate bottlenecks, form hypotheses, run tests, and use AI as an assistant. With /goal, the workflow inverts: you define a measurable target and let the agent work. It maps the repo, generates hypotheses, changes code, runs tests, measures, and iterates until the needle moves. This is solving in computational mode: metric-guided search, in a loop, until a viable solution appears. But here is the important part: the agent’s output should not be treated as final code. An agent exploring autonomously tends to leave traces: hacks, debug prints, failed hypothesis logic. Merging that directly pollutes the codebase. The ideal flow: 1. Treat /goal as a scrappy branch. 2. When it reaches the target, inspect what it discovered. 3. Distill the learning: real bottleneck, working approach, right path. 4. Then write a clean, well-architected PR. The point is not to merge the exploration. It is to learn from it.