Agents in 2026: what actually shipped

A short, opinionated review of the agent patterns that actually shipped this year — and the ones that quietly died.

What shipped

Tool-using single-shot agents. The boring win: a single model call with a tool schema, run inside a small harness that retries on bad JSON. Most “AI features” inside SaaS products are this, dressed up.

Long-running background workers. Email triage, PR review, deal-room research. They run while you sleep and surface a digest. Token costs are amortized over real work output, so the unit economics finally make sense.

Browser-controlled scrapers as research tools. Not “browse the web for me” — that’s still flaky — but “extract this specific field from these 80 listings” is solid.

What didn’t

Multi-agent committees. A planner agent talking to three specialist agents about the same task burned tokens without producing better outputs than a single capable model with the right tools. The “society of mind” framing was beautiful and mostly didn’t pan out for tasks under ~10 minutes.

Autonomy theater. Agents that “think” for 20 seconds and then return a one-line response. Users couldn’t tell whether the latency was buying anything; trust eroded fast.

The pattern underneath

The wins all share a shape: the agent does narrow work the user already understood how to evaluate. Triage, review, extraction. The losses came when the agent was asked to also define the task.

This shouldn’t be surprising — it’s the same lesson every wave of automation has taught — but it took the 2025 hype cycle to make it land.