Generation is supposed to be messy
The creative loop I use is generate, reflect, curate. It is a spiral, not a conveyor belt. I will build several versions of the same idea because each version answers a different question. Agency, DoubleAgency, and Voxelbox overlap for that reason. The overlap is the experiment.
Agents make this kind of generation dramatically cheaper. That is useful, but it also makes the failure mode cheaper: we can produce more plausible things than anyone has time to understand. A learning system cannot treat every output as progress. Generation creates candidates. Nothing more.
Reflection needs the outcome
Reflection is where an execution becomes evidence. What did the agent attempt? What happened afterward? Where did I intervene? Did I change the result because the agent lacked context, chose the wrong tool, misunderstood a policy, or had no way to evaluate its own work?
The distinction matters. “The human edited the answer” is not a lesson. The diff is only the beginning. A useful reflection retains the situation that produced the correction and a theory about why the correction was necessary. Otherwise the system memorizes wording and misses the judgment.
Curation is the trust boundary
ACE systems tend to sound automatic: run the agent, reflect on the result, add the new strategy to a playbook. The dangerous word is add. Unfiltered reflection produces an ever-growing context full of duplicated advice, local exceptions, and confident lessons learned from one lucky outcome. (This is how a system gets older without getting wiser.)
Curation is the quality gate. Is the lesson meaningful? Have we seen the pattern more than once? Does it contradict something already trusted? Is it local to this space, useful to one agent, or general enough to promote into a shared skill? The system can propose the answer. Promotion changes future behavior, so the operator owns the decision.
The next run should start further ahead
Voxelbox records events, outcomes, corrections, lessons, and the provenance connecting them. A confirmed lesson can become a skill. A skill that survives repeated use can harden into a guardrail or move into code. The form changes as confidence increases.
That is the point of the spiral. We generate broadly, inspect what reality did to the idea, and keep only the part that earned another turn. Better context reduces repeated intervention. Better evaluations make reflection less subjective. Better curation keeps the whole thing from collapsing under its own accumulated wisdom.
The next run should start with fewer avoidable mistakes.