The QA Agent Pattern

This is the most underused pattern in Co-Work. And it is — I want to be direct about this — the single most important pattern in this entire course. If you take nothing else from these eight lessons, take this: always pair a creation agent with a QA agent.

The reasoning is simple and it mirrors how every real team operates. Employee 1 creates the deliverable. Employee 2 reviews it before it goes to the stakeholder. Nobody — nobody — presents a first draft to the CEO. Nobody ships the first version of a contract to the client. There is always a review step. Always. The review step is not overhead. The review step is where amateur work becomes professional output.

In Co-Work, this pattern is trivially easy to implement. You spin up a creation agent — "build a 12-slide PowerPoint on our Q1 results" — and simultaneously spin up a QA agent — "review the PowerPoint for text overlap issues, color consistency, logical flow between slides, factual accuracy against the data I provided, and overall presentation quality."

The creation agent builds. The QA agent reviews. Both can run in the same workflow. The QA agent catches what the creation agent missed — text that overflows its bounding box, colors that do not match the brand palette, a chart that contradicts the narrative two slides earlier, a slide that jumps from topic A to topic C without covering topic B.

But here is where it gets genuinely interesting. When the QA agent finds issues — and it will find issues, because first drafts always have issues — you can roast it constructively. "You did a good job catching the text overflow, but you missed the inconsistent header sizing on slides 4 and 7, and the transition from the revenue slide to the forecast slide needs a bridge — the audience will not follow the logic jump." Feed those corrections back.

Now comes the compounding step. After you have corrected the QA agent a couple of times, tell Claude: "Create a solidified skill of how to QA a PowerPoint and what to look for, now that I have corrected you on what you missed." That skill — that specific, refined, battle-tested set of QA criteria — persists. The next time you create a PowerPoint, the QA agent starts from a higher baseline. Your corrections compound into permanent improvement.

The rule is simple and it applies to everything, not just PowerPoint. If you do something more than twice, it deserves to be a skill. And you want to do it the same way every time. QA for PowerPoint. QA for email campaigns. QA for contract language. QA for data analysis. Each one is a separate skill, each one improves through feedback, and each one makes the next iteration faster and more reliable.

The QA agent pattern is also your defense against the most seductive trap in AI workflows: "the first draft looks good enough." The first draft always looks good enough when you are the one who prompted it. You have seen the sausage being made. You fill in the gaps subconsciously. The QA agent has not seen the sausage being made. It comes to the deliverable with fresh context — a fresh 200,000-token window, in fact — and evaluates it as a stakeholder would. That objectivity is worth more than the extra two minutes of compute time.