Don't Lose Your Identity Working With AI Agents (ft. Addy Osmani)

TL;DR

• 1

  Agent productivity gets dangerous fast without identity and governance — Osmani says the jump from “babysitting” personal agents to using them seriously requires knowing which sub-agent is doing what, limiting blast radius, and protecting credentials, API keys, and customer data.

• 2

  Harnesses feel magical when they can infer the path, not just follow instructions — One of Osmani’s standout examples was using Claude Code to process EDM WAV files, figure out where each track should begin a natural fade-out, install the needed tools, and complete the job with minimal hand-holding.

• 3

  You do not need to chase every new harness — With Claude Code, Codex, anti-gravity, OpenCode, and others changing weekly, Osmani recommends treating experimentation time as an “innovation budget” and sticking with tools that already deliver value.

• 4

  The biggest risk is cognitive debt turning into cognitive surrender — As AI generates more code, engineers can lose muscle memory for how things work and eventually stop thinking critically, blindly shipping agent output and assuming the agent can fix any mistakes later.

• 5

  Verification is becoming the real bottleneck — Osmani argues that generation is now easy, but trust requires codifying what “good” looks like through tests, user journeys, visual regression checks, static analysis, and other engineering hygiene.

• 6

  More agents do not create more of you — Even if background agents can handle isolated tasks, Osmani says cognitive bandwidth does not parallelize, so developers need to be deliberate about what to delegate and what still deserves close human attention.