Approve agent work

The "human in the loop" moment. Don't skip it.

The 30-second version

When a workroom reaches completed, open the deliverable preview. Scan section by section. If it's right, click Export. If it's mostly right, click Request revision and pick a preset. If the whole structure is wrong, Replan instead of revising.

Why this matters

The Composer produces a confident-sounding draft regardless of whether it actually understood the brief. A two-minute review catches the kind of errors — wrong client name, hallucinated metric, off-brand tone — that would otherwise show up in front of the client.

Approval isn't a button you press. It's the choice between Export, Revise, and Replan.

How it works

Open the deliverable

When the workroom shows completed, the deliverable section appears on the workroom detail page. You'll see:

• The composed document (rendered Markdown, scrollable).
• Per-task output from each agent (collapsible).
• Activity log on the right (every state change, who triggered what).

[SCREENSHOT: workroom completed view with composed deliverable and task list]

Read it like a junior's first draft

Scan in this order — it's the order most issues show up:

1. Headings. Are the sections the ones the brief asked for? If not, you have a structural problem (use Replan).
2. Numbers. Any pricing, timelines, audience sizes — verify against the brief. The Composer will sometimes invent plausible numbers. Cross-check.
3. Names. Client name, product name, your agency name. Hallucinated names happen.
4. Tone. Does it sound like your agency? If not, your Style Memory needs more samples or the banned-phrase list needs more items.
5. Specific claims. Anything that sounds like an external fact (case study mention, regulation reference) — verify.

Three decisions

• Export. The deliverable is good enough to send. Click Export, pick .docx or .pdf. See export deliverable.

• Request revision. The structure is right but a section needs work. Pick a preset (tighten, change_tone, add_section, remove_section, client_feedback, etc.) and write a one-line instruction. See request revisions.

• Replan. The structure itself is wrong (missing whole section, wrong agent did the work, dependencies were off). Click Replan — this discards non-completed tasks and re-invokes the Planner.

What approval is not

There is no separate "approve" button that locks the deliverable. A completed workroom can still be revised any number of times. Exporting doesn't lock it either — you can revise after exporting and re-export with a -v2 suffix.

There is also no per-task approval flow surfaced in the default workroom UI. If you build custom agents that include approval task types in their plans, those tasks will pause and ask for an explicit grant — but that's an agent-builder feature, not the default review flow.

Common pitfalls

• Approving on the basis of length. The Composer can produce 4 polished pages of nonsense. Read the headings and numbers, not the word count.
• Revising structural problems. If the section list is wrong, Replan. Trying to fix structure with a long custom_instruction revision often wastes credits.
• Treating the activity log as optional. It tells you which agent did which task. If the wrong agent ran, the activity log is the fastest way to find out.
• Skipping the read because the workroom shows green. "Completed" means the state machine reached the end, not that the deliverable is correct.

What's next

• Request revisions — preset-by-preset guide.
• Export deliverable — once you're ready to ship.
• Approval and revisions — the model behind the flow.