Six agents walked into a productivity test, and only one understood what work actually looks like.
The Summary
- Wired tested six major AI agents on basic task completion: Meta's Muse, Anthropic's Claude Duet, and four others through real-world workflows
- The winner wasn't the one with the biggest model or the flashiest demo, it was the one that understood context handoffs between tasks
- Task completion rates varied by 40% between the best and worst performers on identical assignments
The Signal
The test wasn't about benchmarks or synthetic problems. Wired ran six agents through actual knowledge work: schedule a meeting across three calendars, research a topic and draft a brief, order lunch for a team with dietary restrictions, follow up on three email threads without dropping context. The kind of work that eats two hours of your day before you get to the work that matters.
The results exposed a gap between what AI labs ship and what actually runs in production. Half the agents failed at context persistence. They'd nail the research task, then forget key details when drafting the brief. Or they'd schedule the meeting but lose track of which participant had the conflict.
"The best agent wasn't the smartest. It was the one that remembered what you told it three steps ago."
Claude Duet came out on top, not because it had the most parameters, but because Anthropic built it to handle task chains. When you asked it to research competitors and then draft an email to your team about findings, it carried forward the details that mattered. Meta's Muse was close behind, strong on consumer-friendly tasks but weaker when work required crossing between different tools and contexts.
Key gaps across all six agents:
- None handled ambiguous instructions well, they all demanded clarity humans rarely give each other
- Integration quality mattered more than model capability, the agent with the best calendar API won the scheduling test
- Error recovery was universally poor, when an agent hit a wall it stopped rather than finding a workaround
The Implication
This test matters because it shows where the agent economy is actually building versus where the hype lives. If half the agents on the market can't chain three tasks together without losing context, we're still early. The companies winning Web4 won't be the ones with the biggest models. They'll be the ones who solve handoffs, memory, and tool integration.
For anyone buying or building with agents right now: test them on your actual workflow before you commit. The demo that works in a product video falls apart when it has to talk to your calendar, email, and Slack at the same time.