Eighteen months ago the conversation about agentic AI was about capability. Could agents plan, could they call tools reliably, could they recover from errors. By mid-2026 the capability questions have mostly settled — the answer is yes, with caveats. The interesting conversations are now about which deployment patterns have actually stuck in production.

Pattern one: narrow agents with hard constraints

The agents that are running reliably in production at scale are not the open-ended ones. They are narrow agents with a constrained action space. Customer service deflection, internal data lookups, structured workflow automation. The agent has a small set of tools, a clear objective, and explicit guardrails on what it cannot do.

The pattern is not glamorous. It is durable.

Pattern two: human approval at consequential steps

The other pattern that has stuck is hybrid execution. The agent runs autonomously until it reaches a step with real consequences — moving money, sending external communication, modifying production data — and pauses for human approval.

Teams that deployed fully autonomous agents in 2025 have walked back to this pattern. The trust required for full autonomy in consequential domains is not there yet, and the failure modes are too expensive.

Pattern three: routing rather than reasoning

A small classifier model decides which agent or which tool should handle each request. The router is fast and cheap. The downstream agent or tool is task-specific. The routing pattern beats the monolithic-agent pattern on cost, latency, and reliability.

What has not worked

Long-horizon planning agents have been the consistent disappointment. The capability is real for short horizons. As planning horizons stretch beyond a few steps the error compounding becomes operationally unworkable.

Multi-agent debate and self-critique patterns have produced research papers but few production deployments. The latency and cost overhead is hard to justify outside narrow use cases.

Where the genuine value is showing up

Customer service deflection at 60-80% rates with sustained CSAT. Internal knowledge agents that turn ticket resolution from 20 minutes to 2 minutes. Structured data extraction at near-human accuracy across millions of documents. The boring categories produce the value.

For organisations thinking through their first production deployment, AI agent development firms that have shipped these patterns are more useful than those leading with capability demos. The implementation craft now matters more than the model selection.