Agent architecture
An agent is not magic prompt glue. The reliable version has a narrow workflow, explicit tool permissions, durable state, retries, human approvals, evaluation, observability, and a fallback route when the model or tool fails.
Last updated September 17, 2026. Agent frameworks change quickly, but the safety and reliability checklist stays useful.
Blueprint
Best beginner scope: pick one repeatable task, one or two tools, one approval step, and one success metric. Avoid open-ended "do anything" agents until your logs and failure handling are boring.
Cost and protocol discipline: OpenAI's Codex team calls parallel sub-agent sprawl a "coordination tax" - more than two sub-agents usually burns tokens without quality gains. Connect tools through MCP (Model Context Protocol) servers so permissions, audit logs, and tool contracts stay centralized across Claude, ChatGPT, Cursor, and VS Code clients.
Test whether the model follows schemas, handles retries, and asks for missing information before using tools. Nex-N2.5 (free tier on OpenRouter) and DeepSeek V4.1 Flash are strong agentic options; Mercury 2.5 fits latency-critical loops.
Every tool needs input schema, permissions, rate limits, error messages, and audit logs.
Track tasks, steps, decisions, tool outputs, approvals, and retries in a durable store.
Use scenario tests, tool error rates, human review outcomes, latency, and cost instead of vibes.
Guardrails
Most useful agents are constrained operators. They know which tools they can call, when to ask for help, when to stop, and what evidence must appear before action.
People also ask
A narrow workflow that reads one input, calls one safe tool, asks for approval when needed, and writes one auditable output is often better than an open-ended autonomous agent.
They need durable state more than vague memory. Store tasks, tool outputs, decisions, user preferences, and approvals in structured data.
Use a stateful framework when the workflow has branching, retries, approvals, and long-running jobs. For one-step tools, direct API calls may be enough.