An agent can produce a valid response without completing the requested work. It can also leave behind a useful artifact while the conversation ends or the next step remains disconnected. To continue safely, a successor needs to know what the artifact is for, which request it answers, and which current assumptions support it.
The three essays below examine different parts of that problem. A design note distinguishes products built from the same tools. A real batch-processing failure shows why success belongs after identity checks. A reading of Flowness's Work design asks how the resulting state could remain queryable beyond one executor's memory.
The same three tools, three different products ↗
Choose whether an air-quality project first owes its user a transformed record, an investigable comparison, or an explanation that stands alone.
Commit success after checking identity ↗
Follow an item whose case identifier changed and learn why both new calls and historical success receipts need the same acceptance boundary.
Give work an address beyond the conversation ↗
Inspect a scheduler handover where the tests pass an obsolete requirement, and see what a current work view must preserve.
A useful check for an existing workflow is to interrupt it deliberately in a reversible test. Can another executor identify the latest objective, the actual files and results, the next necessary action, and what still awaits real use or acceptance? Count the work this saves against the effort of maintaining the records.