REAL EXPERIMENT

How We Taught Our Codex Workflow Anthropic's Principles

Level: intermediateReading time: 10 minutesResult: a verifiable AI-agent workflow

We did not retrain the model or change its weights. We studied Anthropic's courses, tested the ideas on Hermes, and rebuilt the way our own system works with Codex.

From a model to a working system

The most useful lesson was practical: an agent is not reliable just because it can answer. It needs boundaries, context, tools, verification, and a way to stop. The model supplies reasoning. The surrounding system supplies the operating discipline.

What changed in our workflow

Complex work now starts with a plan. A delegation records the expected deliverable, inputs, constraints, completion criteria, review point, allowed tools, and execution limits. This makes “done” observable instead of leaving it to the agent's intuition.

We use instruction files for conventions, but hard rules live in hooks and API gates. An agent may draft an email automatically. It may not send that email without owner approval. The same separation applies to publishing, deleting files, changing credentials, and financial actions.

Verification is more important than a polished answer

We added reusable Skills and evaluation cases for recurring workflows. The check is not only whether the final text looks good. We also check whether the required steps ran, whether the agent weakened a test to make it pass, whether the result matches the source material, and whether unresolved uncertainty is visible.

For material changes, a second reviewer gets a clean context. The agent that produced the result should not be the only authority declaring that result correct.

The email pipeline became a practical test

Hermes saves original messages, extracts dates and amounts, removes duplicates by mailbox identity and content digest, separates useful work from noise, and keeps external actions behind approval. Some categories are deliberately silent: enforcement notifications and procurement invitations remain in the knowledge base but do not create a Telegram message for every email.

When Codex reached its usage limit, we added a persistent backoff. Instead of producing a storm of identical errors, the system preserves incoming mail and waits for access to return. The queue can resume without losing the source material.

Why we did not enable everything

Anthropic's platform courses also made the cost of unnecessary integrations clear. We did not automatically enable every API, connector, managed agent, or third-party plugin. Each new capability must have a measurable benefit, limited permissions, a cost boundary, a verification method, and a rollback path.

The result

We did not turn Codex into a different model. We taught our working system to use it responsibly: plan before execution, keep contexts bounded, route simple work to cheaper paths, verify outputs independently, and tell the owner what remains unconfirmed.

← All practical articles

We publish what we test in our own lab. Practical AI agents, automation, and action controls for real business workflows. Discuss a project →