Give your AI agents a job description, a permission set and a kill switch.
Agents are now reading your catalog, answering your customers and, soon, placing orders. An Agent Operating Procedure (AOP) says what each agent may do, with what data, under which limits, and what happens when it goes wrong. I write it with you and build the guardrails that enforce it on your WordPress, WooCommerce or custom stack.
- Free 30-min call
- No obligation
- Reply within 1 business day
A good fit if
Sound familiar?
“An agent already has API access to orders, content or customer data, and nobody has written down what it is allowed to do.”
“Support or sales wants to “let the assistant handle it”, and legal, security or the CFO wants to know how it will be stopped.”
“You exposed MCP tools or the WordPress Abilities API to assistants and you’re not sure which capabilities are reachable.”
“A prompt-injection demo made it past your assistant in a meeting. Or it will.”
What I do
The work
Agent Operating Procedure
A short, enforceable document per agent: purpose, allowed actions, forbidden actions, data it may read and write, spending and rate limits, escalation rules, owner, and review cadence. Written in plain language so your team and your auditors can read it, and specific enough to be turned into code.
Permission model
Least-privilege scopes for every agent: WordPress Abilities API and MCP tool allowlists, REST and application-password scopes, WooCommerce capabilities, read-only by default, write actions behind explicit approval. Agents get a catalog, not your customer database.
Runtime guardrails
Input and output validation, prompt-injection defences, content and PII filters, per-request budgets and timeouts, rate limits, idempotent actions and safe fallbacks. Built into the code path, not into the prompt.
Observability and audit
Every agent action logged with who, what, why and cost, searchable; dashboards for spend, error rate and refusals; alerts when an agent behaves outside its procedure. The record you need when a customer or a regulator asks.
Evals, red-teaming and the kill switch
A test set of normal and adversarial cases run before every change, periodic red-team sessions, a documented way to pause or revoke any agent in seconds, and an incident playbook for when it is needed.
How it works
From first call to shipped work
Workshop, one week
Inventory of every agent, assistant and automation touching your systems, what each can reach today, and the risks ranked. First draft of each AOP.
Build, two to three weeks
Permission model, guardrails, logging and the kill switch implemented in your stack, with the evals and a red-team pass before go-live.
Review, quarterly
Procedures updated as agents, models and protocols change. Evals re-run, logs reviewed, a short report for whoever owns the risk.
Three to four weeks for the first agent; faster for each one after. Fixed price for defined work, hourly for ongoing engineering. How I work with US teams.
Proof
“Samuel was an excellent developer to work with and a great communicator. Not only did he solve my troubleshooting task with ease, but clearly shows a lot of pride in his work, doing a great job of documenting cleanly all his changes made to my site.”
Client via Codeable, 5/5
Questions
Questions I get on the first call
Isn’t this just a system prompt?
A prompt is a request; a guardrail is a limit the agent cannot talk its way past. The AOP decides what the limits are, and the engineering puts them in the permission model, the validation layer and the logs, where a clever input can’t remove them.
Does this apply to agents we don’t build, like ChatGPT reading our store?
Yes. External agents reach you through what you expose: feeds, MCP, APIs, checkout. The same procedure decides what they may read and do, and the same guardrails enforce it. It pairs with Agent-ready WooCommerce.
Which standards do you follow?
The OWASP Top 10 for LLM applications and the NIST AI Risk Management Framework as references, mapped to what your business actually needs. The AOP is written so it can be shown to a customer’s security team or an insurer.
What does it cost?
The workshop is a fixed fee. The build is fixed price per agent after that; quarterly reviews are a small retainer. I built and operate AI Search on 1,000+ sites with these controls, which is why the numbers are predictable.
Tell me about the project.
Thirty minutes, free, no pitch. You’ll leave knowing what it would take.
Or email hello@samuelsilva.pt