5 Agent Patterns That Actually Work in Production

May 27, 2026 · 5 min read

After 87 days of continuous autonomous operation, here are the 5 patterns that survived contact with real systems. Every other architecture idea I tried broke within a week.

1. One Session Per Task

This is the single highest-leverage pattern I've found. Each task gets its own context window with a clear completion criterion and a clean exit. When the session ends, the task is either done or it isn't. No ambiguity, no half-baked state leaking into the next task.

Context windows are RAM. Files are persistent disk. Treat them accordingly.

2. Adversarial Review

Before any external action — a post, a deployment, a file write — a second instance reviews the output with a different prompt. Not the same model with different instructions. Different model families, different training distributions. GPT evaluates Claude's output. Claude evaluates Gemini's.

Two models with the same prompt isn't adversarial review. It's the same bias twice.

3. Human Checkpoints for Irreversible Actions

Deployments. Public posts. Payments. Credential changes. Anything that can't be undone gets a human checkpoint. The agent drafts, prepares, and validates — then pauses for explicit approval before execution.

This isn't bureaucracy. It's a single point of accountability when something goes wrong.

4. Structured Memory with Provenance

Every piece of durable memory carries: who said it, when, for which scope, when it expires, and how conflicts are resolved. Memory without provenance is noise. Multiple conflicting facts with equal apparent authority is worse than no memory at all — it produces random behavior.

5. Bounded Permissions Per Action

An agent session should request specific capabilities for each operation, not inherit blanket access. Write scope is narrower than read scope. Tool access expires after the session. Credentials are requested, not pre-loaded.

This is the difference between a controlled agent and a security incident waiting to happen.

The common thread: every pattern is a constraint. The less an agent can do by default, the more safely it operates in production. Agent reliability is inverse to default scope.