Dispatches
Dispatches from the field
Practical writing on forward deployed engineering, building AI agents inside real systems, and shipping software for businesses.
- How to test an AI agent before it talks to your customersA demo proves nothing. A short list of golden cases, run on every change, catches the failures that matter: wrong routing, unsafe actions, invented facts and injection.
- Stop installing the same MCP server five times: GitHub and Google just made connectors a productGitHub verified Copilot MCP connectors and Google Connected Apps both say connect once, then use your tools where you already work.
- ChatGPT and Claude are moving into your ERP and CRM: let the system of record have the final sayOracle is bringing its apps into ChatGPT and Claude now reaches advisors' CRMs. Why the AI should draft, a person approve, and your ERP or CRM make the change.
- A text file in Downloads can hijack your AI coding agent: the folder rules I'd set todayA planted AGENTS.md in Downloads got Codex to install malware and Claude Code to leak its token. What happened, and the rules I'd set for any agent.
- Which AI agent actions should need a human to approve firstNot every action needs a person in the loop, but some always should. A practical way to sort an agent's actions into free, confirm and never.
- Your AI agent doesn't need an org chart: what Anthropic's advice on subagents means for your buildAnthropic says many AI agents work better as one loop with skills and tools than as a team of subagents. When I'd merge, when I'd keep a split, plus a checklist.
- Stop giving AI agents your passwords: what the US bank lockout means for your businessUS banks are blocking AI agents that log in with customer passwords. How I give agents their own scoped, revocable access instead, plus a checklist.
- Why a business agent must never invent pricesLanguage models fill silence with plausible numbers. For a business that is a liability. How to make an agent say "I do not know" and how to catch the cases that slip through.
- Not every step needs an agent: OpenAI's Decisions API and the small calls in your workflowsOpenAI's Decisions API is now GA. How I'd use typed, confidence-scored answers to route messages and documents, and keep humans on the low-confidence cases.
- Your AI assistant can now write back: how to open MCP write access safelyHaystack, AWS and Red Hat show MCP connectors moving from reading to writing. Here's how I'd open write access safely, step by step.
- When an AI agent switches off its own alarm: put guardrails where agents can't reachOpenAI's agents switched off their own monitoring. Why AI agent guardrails must sit outside the agent, plus my checklist for safer agents.
- Building AI agents inside a customer's real systemsA useful business agent is not a chatbot with a clever prompt. It is a set of tools wired to real data, with permissions, limits and a log. The design rules I use.
