I write about durable software and build open-source tools that keep humans and agents aligned.
Hi, I’m Gitesh. I design software systems and document what holds up as AI agents reshape engineering practice—from delivery workflows to the context they need to do reliable work.
An agent's memory fails in two ways: it rots inside a long session, and it vanishes entirely between them. The fix is not a bigger context window. It's treating every session like onboarding a new teammate — and keeping the onboarding doc in the repo.
Sep 23, 2026 — Eleven weeks of building with FDF: how it changed my job, and why Fable 5.1 and Opus 5.5 made written specifications matter more, not less.
Sep 23, 2026 — GPT-6 Sol and Luna lead with cost per task instead of raw scores, and that moves the expensive part of agentic engineering from writing code to checking it.
Aug 28, 2026 — OpenAI's Hugging Face postmortem, Claude Code's new restricted mode, and Okta's Agent SSO shipped in the same week, and each is an answer to the same question: what stops an agent that cannot stop itself.
Aug 28, 2026 — MCP deleted sessions from the protocol, Opus 5 started writing corrections into its own memory, and a dozen frontier models shipped in three weeks. Read together, those are the same announcement.