An idempotency key contract for AI tool calls
A practical contract for making mutating AI tool calls safe across retries, resumes, parallel execution, and MCP integrations.
Agents and developer tooling
AI Systems
A builder's view of AI systems: how agents are evaluated, constrained, observed, and integrated into real software workflows.
20 articles
A sandbox manifest contract for coding agents
Treat coding-agent sandboxes as explicit contracts: files in, paths granted, secrets isolated, state persisted, and traces tied to review.
A widget contract for ChatGPT Apps SDK launches
ChatGPT apps need a small contract between MCP tools, widget state, and permission prompts. This checklist keeps rich UI useful and reviewable.
A budgeted model router for GPT-5.6 workloads
A practical playbook for routing AI requests across model tiers, budgets, and fallbacks without turning provider choice into application sprawl.
A state budget for production AI agents
A practical checklist for deciding what an AI agent should remember, checkpoint, compact, replay, and forget before it handles real work.
Claude Opus 5 is here: what developers should test before upgrading
Anthropic's Opus 5 pairs near-Fable capability with Opus pricing. Here is the practical migration checklist I would use before moving a production AI workflow.
Human handoff gates for long-running AI agents
Long-running AI agents need explicit handoff gates around authority, evidence, cost, and irreversible actions before teams can trust them in production.
A telemetry contract for production AI agents
Production agents need a small trace contract before they need a new dashboard. Start with workflow, model, tool, cost, and privacy signals.
Treat agent instruction files like production code
Agent instruction files now influence PR quality, token spend, and safety. Review them like code, with scope, ownership, tests, and expiry.
A governance checklist for remote MCP servers
Remote MCP servers make agent integrations portable, but the production work is in token boundaries, tool approvals, logs, and change control.
GPT-5.6 Is a Family, Not a Model: A Field Guide to Sol, Terra, Luna, Max, and Ultra
OpenAI's newest release is a three-model family with new reasoning and agent modes. Here is what Sol, Terra, Luna, max, and ultra actually mean—and how I would choose among them.
The Agent Supply Chain: 7 Layers Founders Are Selling to AI Agents
A field guide to seven layers serving AI agents, from identity and payments to sandboxes, memory, and compute infrastructure.
7 Rules for Rolling Out CLI Coding Agents After Microsoft's Study
A July 2026 Microsoft study puts real numbers behind CLI coding agents. The practical lesson is not just to buy seats, but to design the rollout like an engineering system.
A field guide to the agent economy: the stack being built to sell to AI agents
AI agents are becoming customers, and a whole startup stack is forming around them. Here is a layer-by-layer map of the payments, identity, memory, and compute plays worth watching.
A Q&A on Google Vids, Veo 3.1, and everyday AI video
A practical Q&A on Google Vids and Veo 3.1: where AI-generated clips and avatars fit everyday presentation work, and where they fall short.
An anatomy of Vercel Sandbox for agentic code execution
A practical breakdown of Vercel Sandbox for AI-generated code: execution boundaries, runtimes, lifecycle, costs, and integration decisions.
7 myths about AI agents after the Codex adoption data
A practical myth-vs-reality read on what the June 2026 Codex adoption paper says about agents, developer workflows, and the gap between demos and durable habits.
MAI-Code-1-Flash vs Claude Haiku 4.5: Microsoft's first OpenAI-free coding model, head to head
Microsoft shipped MAI-Code-1-Flash into GitHub Copilot at Build 2026, its first coding model trained without OpenAI tech. I put its claims against Claude Haiku 4.5, round by round.
Agent infrastructure is the startup bet worth making in 2026
The flashy agent products get the headlines, but the durable startup opportunity in 2026 is the plumbing underneath: identity, payments, memory, and sandboxes.
A small, honest eval harness beats vibes
Most teams ship LLM features on gut feel. A tiny, boring evaluation script catches regressions before your users do.