Weekly Tasks are now ready for agents to carry out
Every Task now ships with the target, evidence, deliverable, and completion rule an agent needs, and MCP clients can read a batch, execute it, and record the result.
What changed
- Each Task carries a structured execution package: exact target, mechanism, deliverable template, execution steps, evidence refs, completion artifact, and watch plan.
- Tasks are labelled agent_ready or needs_review, so an agent knows which work it can pick up without a human first.
- MCP clients can now list past Task batches and read a historical batch with its receipts, evidence, and watch state.
A Task now carries everything needed to do it
A Task used to be a recommendation with its reasoning attached. It is now a structured package. Alongside the exact target and the mechanism, each one carries the destination and the rule that validates it, a deliverable template with its required sections, the execution steps and the decision boundary on each operation, the inputs required, the evidence it rests on, the artifact that counts as completion, and the window to watch afterwards.
It also carries what was deliberately left out. Evidence exclusions are recorded with a reason, so whoever picks the work up can see which signals were considered and rejected rather than guessing what the absence means.
Tasks say whether they are safe to automate
Not every Task should go straight to an agent. Each one is now scored against its response-backed evidence and labelled either agent_ready or needs_review, with the reasons attached when it is the latter.
That label is the difference between an agent that helps and an agent that confidently does the wrong thing on your website.
The full loop runs over MCP
A connected MCP client can generate a batch, read the current Tasks, pull a Task's execution package and evidence summary, and record the completion when the work is done.
New in this release: listing persisted Task batches and reading an exact historical batch, complete with completion notes and receipts, evidence, execution summaries, and watch state. Reads stay brand-scoped, entitlement-checked, bounded, and cursor-paginated, so a client sees one brand's work and nothing else.
The quality work that made this safe
Handing Tasks to an agent is only an improvement if the Tasks are good. Several weeks went into making that true. Batches now contain interventions only, each with a concrete URL deliverable and an observation plan, rather than prerequisite work like readiness checks and inventories. Every valid request releases exactly three evidence-backed Tasks with distinct mechanisms.
The most important change is the least visible. Completion receipts are now a hard eligibility gate, checked before ranking. We found this on our own account: a batch had reused completed outcomes across all three active prompt cohorts, with one candidate changing only the URL slug while keeping the same cohort and mechanism. Generation now excludes completed equivalents and fails closed with fewer than three Tasks rather than recycling or inventing work.