Dumping every workstream's context into each agent is a poor way to delegate. On June 17, 2026, I objected to exactly that idea. I want agents collaborating, and I still want to know which job each one received. Mixing everything together as a precaution pushes the decision about relevance onto the next executor.
The following day, Codex Record & Replay was announced with the proposal of generating reusable skills from demonstrations. To me, that raises a concrete choice: what should travel with a capability when it leaves the first execution? Repeating a procedure does not necessarily require carrying every conversation that led to discovering it.
Consider checking a delivery as a hypothetical example. The checking procedure could apply to other tasks. The history explaining one specific decision belongs to that case. If they become an inseparable package, the next agent has to discover which passage is a general instruction and which only made sense in the previous job. Then we complain that it got confused. We helped.
I want to separate the capability description from the material for the current execution. The skill explains when it applies and which criteria guide the work. The task supplies the context needed to use it now. That separation makes it easier to check whether the answer came from the current request or from old information that travelled along without a reason.
History can be necessary. An earlier decision might rule out a solution that looks good when reading only the latest request. In that case, carry the decision and explain its consequence. Cutting context at any cost causes rework too: an agent repeats a settled investigation or proposes something incompatible with a known constraint.
The boundary I propose is straightforward to enforce: each relevant piece needs a reason to accompany the task. Missing information should become a concrete question. Material that changes no decision can be removed in a later comparison while keeping the expected result constant. That lets the work reveal the boundary instead of choosing a package size through superstition.
I will organize context handoffs around the delegated objective and leave history available when there is a reason to consult it. A skill should carry reusable knowledge; each execution gets what it needs to act and check the result. If using a capability requires supplying every conversation from every workstream, the boundary is still poorly defined. More material does not relieve me of the job of defining it.