What Changes When an Agent Ships the Finished Work

ChatGPT Work, which OpenAI launched alongside GPT-5.6 on July 9, 2026, pulls context from connected apps and files and produces finished artifacts — sheets, slides, documents, web apps — by breaking a goal into steps and working through them independently for hours. Unlike an action agent that ends after one tool call, an outcome agent is built so a human doesn't have to watch every step from kickoff to final delivery, and that alone widens the trust boundary it operates inside.

Why Tool-Call Logs Stop Being Enough

A conventional tool-call log records which function ran with which arguments, but it doesn't tell you which sentence in the final document traces back to which fact pulled from Slack, Notion, or Google Drive over a multi-hour session — unless you log that mapping separately. GPT-5.6 ships in three sizes, Sol, Terra, and Luna, priced at $5, $2.50, and $1 per million input tokens respectively and routed by task difficulty; because citation behavior can shift with the size handling a given step, a sentence-to-source mapping log is the only way to verify an artifact after the fact.

The Weight of Work That Leaves the Building

A coding agent's output passes through a PR review gate before it ships. A dashboard or executive report an outcome agent assembles often has no equivalent gate — it goes straight to stakeholders. The fact that a deliverable no human touched mid-stream now has a direct path to an external audience is the question this release actually poses to QA teams.

From Design to Rollout: An Outcome-Agent Operations Checklist

Start by splitting review criteria numerically by artifact type. Set the auto-approval rate for externally shared artifacts to 0% — human sign-off required every time — and target a human-intervention rate of 15% or less for internal drafts. Cap simultaneous connected apps per session at four; beyond that, least-privilege gets hard to hold, so pre-approve a per-task-type app whitelist and route any request outside it through separate approval. Cap task duration too: past six hours, force an automatic checkpoint to confirm the agent hasn't drifted off the original goal.

Failures cluster into three shapes: a stale fact pulled from one of several connected apps making it into the final document unchanged; context drift over a long session, where the six-hour output quietly diverges from the original instruction; and an earlier file version getting overwritten with no review and no way back.

Saving an artifact snapshot alongside the tool-call history at every checkpoint lets you roll back to that point the moment something looks off — the same principle covered in cross-device agent session handoff for resumable checkpoints. Detect drift by periodically re-measuring the semantic distance between the original instruction and current progress, and auto-halt at the threshold so the agent doesn't advance to the next stage without a human check.

Require a pre-publish gate for any artifact meant for external sharing, with citation completeness — a source app or document link behind every sentence — as a pass condition. Log the apps accessed, the call order, the per-sentence source mapping, and total session duration as mandatory fields.

Any session that crosses PII-bearing sources — customer emails, contracts — should run automatic masking before artifact generation, with a safe-abort path that halts generation outright if masking fails.

Weekly, tally the failure types behind gate rejections to see which app combinations or task types produce the highest rejection rates. Drop the worst-performing combinations from the whitelist, or tighten next week's review bar for them.

Takeaways at a Glance

As outcome agents stretch into hours of unattended work, review has to be designed around a duration, not a single checkpoint. Lock in a 0% auto-approval rate for externally shared output, a four-app connection cap, six-hour checkpoints, and sentence-to-source mapping logs first, and the gate keeps working even after the next model size lineup changes.

References

ChatGPT Work with GPT-5.6 — OpenAI

GPT-5.6: Frontier intelligence that scales with your ambition — OpenAI