Notes on AI agents and MCP
Short, plain notes about building AI agents and MCP servers that work in real businesses: what works, what breaks, and why.
June 5, 2026
"Which part do we agentize first?" is the wrong first question
The whole market has moved from 'are agents real?' to 'which part of my company gets agentized first?' — IT support, sales, reconciliations. It feels like the smart strategic question. It's the wrong one. Asking where to point the agent skips the two questions that actually decide whether any of it works: what does the agent stand on, and who answers when it's wrong. Here's the order that matters.
- methodology
- business
- agents
June 4, 2026
87% on the benchmark, and it still can't evolve your codebase
The headline says AI 'solves 87% of SWE-bench,' and everyone reads it as 'AI can do software engineering now.' Two problems. The small one: a third of those passes leaked the answer or had weak tests. The fatal one: the benchmark measures one isolated bug fix, not the actual job — evolving a living codebase over weeks. Measure that, and the same models fall from ~73% to ~25%. The benchmark is the demo. Your codebase is production.
- eval
- agents
- methodology
June 4, 2026
Your agents are stateless. That's why they die.
Your agent finishes start-to-finish on your laptop, so you never see the problem. Production is a long, multi-step process on infrastructure that restarts, times out, and dies partway — and your agent kept all its progress in memory. The 2026 'agentic reckoning' is the discovery that the failure isn't the model, it's the runtime. The fix is old and boring: durable execution. Here's the honest version.
- architecture
- agents
- methodology
June 4, 2026
Tokenmaxxing, or Goodhart's law comes for AI
Companies wanted 'AI adoption,' so they measured the easiest proxy — token usage — put it on a leaderboard, and got exactly what they measured: people burning tokens to climb the board, not to do better work. It's a fifty-year-old law eating a brand-new strategy, and now it's expensive twice: you pay for the wasted tokens and you poison the signal you wanted. The fix is old too — measure outcomes, not activity.
- methodology
- business
June 4, 2026
Your org chart can't run agents
Every exec is asking 'is the model good enough yet?' New MIT data says that's the wrong question: 85% of organizations want to be agentic, but 76% admit their own operations can't support it. The blocker isn't model capability — it's that companies are bolting a new kind of worker onto an org chart drawn for humans. An agent has no manager, no career ladder, scoped permissions, and a hallucination rate. Here's the box your chart is missing.
- business
- methodology
- ai-native
June 3, 2026
Most AI agents never reach production
The demo is dazzling. Then the agent never ships. Survey after survey in 2025–26 finds the same cliff: almost everyone has an agent pilot, almost no one has it in production. The reason isn't the model — it's the unglamorous engineering the demo let you skip. Here's what the small minority who actually ship do differently.
- agents
- eval
- methodology