"Agents" was the most over-used word at every 2026 keynote, Anthropic's, OpenAI's, Google's I/O. Demos look magical. Production deployments are messier. After a year shipping AI agents into real apps for real customers, here is the honest field report.
What an "agent" actually is in 2026
An agent is an LLM that can take more than one step, use tools, and decide what to do next. That's it. The buzzword bingo around it ("autonomous", "self-healing", "AGI-adjacent") is mostly marketing.
The three agent patterns that actually ship
- Task agents. Short-lived, goal-bound: "draft this email and send it," "extract these fields and store them." 3–8 tool calls. Reliability is high. This is 80% of what we deploy.
- Workflow agents. Multi-step, branching: "ingest a support ticket, classify, draft, route, follow-up if no response in 24h." Reliability is good with strong guardrails and human checkpoints.
- Coding agents. Claude Code, Google Antigravity, Cursor agent mode. The most successful agent category in the industry by revenue, because the developer is the human-in-the-loop.
What still doesn't work
- Long-horizon autonomy. "Run my entire ops for a week." No. Don't even try.
- Open-ended browsing. Agent demos that browse the web at length are still flaky once you leave the demo's happy path.
- Critical-path autonomy. Anything that touches money, medical, or legal, humans stay in the loop. The agent assists, doesn't act.
The architectural patterns that matter
- Tool boundaries are everything. Narrowing tool surfaces (read-only DB, scoped APIs) lifts reliability dramatically. Wide tool surfaces == hallucinated tool calls.
- Checkpoint & review. Agents that pause at decision points and ask for confirmation outperform fully autonomous variants on enterprise tasks.
- Persistent memory + scratchpad. Long tasks need both, short-term scratchpad for the current task, long-term store for facts that should outlive the session.
- Observability. If you can't replay an agent's decision tree, you can't fix it. Log tool calls, intermediate reasoning, and corrections.
Frameworks worth your time
- Anthropic's Claude Agent SDK. Best-in-class for production. Built on Claude Code's harness.
- Google Antigravity (post-I/O 2026). Strong if you're already in Google Cloud / AI Studio.
- OpenAI Agents SDK. Solid, integrated with the OpenAI ecosystem.
- LangGraph / DSPy. Useful for complex graph-based flows when you need explicit control.
The KPIs we track
- Task completion rate (not just LLM output rate).
- Mean tool calls per task, high counts often signal a fragile prompt.
- Human-correction frequency, the real reliability number.
- Token cost per completed task, the only number finance cares about.
Where to bet in 2026
Coding agents, customer-support agents, structured-extraction agents, and internal-ops agents are paying off now. Avoid betting on fully-autonomous, long-horizon agents until 2027 at earliest.
Want agents in your product done right? See our AI voice assistant case study or talk to our AI team.
