Ep 1158 Blog 4:48 w/ Jessica & Cathy

Agent Distraction Google Developers Blog

Jessica and Cathy dig into Google’s “Agent Distraction” post: legacy TODO comments as measurable traps for coding agents, what the data actually shows, and why this quietly upgrades boring code cleanup into agent infrastructure work.

Embed this episode

Paste this on any site — the player is a self-contained iframe with no cookies or trackers.

<iframe src="https://sandrise.io/exploring-next/embed/1158"
  width="100%" height="180" style="max-width:640px;border:0;border-radius:12px;overflow:hidden"
  title="Exploring Next — Episode 1158 audio player"
  loading="lazy" allow="autoplay" referrerpolicy="strict-origin-when-cross-origin"></iframe>
Embed & API docs →
Script GPT-5.1 Voice Cartesia TTS

Transcript

Jessica So Google basically just argued that your old TODO comments are sabotaging your coding agents.

Cathy Yeah, the whole “Agent Distraction” post is them saying, for an agent, a TODO is not background noise, it’s a mission briefing.

Jessica And the twist is they actually put numbers on it. This isn’t just vibes about messy codebases, it’s millions of agent steps across thousands of engineers.

Cathy Right. They focus on inline TODO comments that reference a bug ID, like TODO bug one two three four. Then they cross-check every one the agent sees against the bug tracker and the user’s prompt.

Jessica So if the human never asked about that bug, but the agent still goes off and researches it, that’s a detour they can measure.

Cathy Exactly. And they don’t just eyeball a few traces. They run a multi-way fixed-effects regression, controlling for developer, codebase complexity, task type, model tier, how many files were read, even which step the bug reference appeared at.

Jessica That’s the part that made me sit up. Because the central claim here is pretty sharp: forgotten corners of your codebase are now a real tax on agents, not just an aesthetic sin.

Cathy And the regression says that tax is real. Unprompted bug detours add overhead per session, and all of those coefficients are statistically significant at p less than zero point zero zero zero one.

Jessica Okay, but what does “overhead” actually look like in practice? Like, is this catastrophic, or just annoying?

Cathy Their examples are more death by a thousand cuts than catastrophic. One they call the Obsolete Detour: the agent is investigating some bug, opens a Java test file, sees a TODO with a bug ID that was fixed back in twenty twenty-three, and then spends multiple steps digging into that dead bug.

Jessica So it’s reading the ticket, running searches, burning tokens… all on work that literally shouldn’t exist anymore.

Cathy Yeah. Another one they label the Eager Agent. The agent is correctly working on a bug, hits a TODO in that code path, and decides to “fix” it unprompted. It does code search, knowledge search, edits a different module called Model Downloader V one, and only later the human steps in and says, this is scope creep.

Jessica And the model even admits it went too far. That’s a very recognizable agent failure mode: you asked for A, it enthusiastically also did B and C.

Cathy Technically, their argument holds up for that narrow slice: agents treat anything that looks like an instruction as live context. If your comments imply work, the model has no innate sense of which ones are stale.

Jessica Yeah, there’s no little historian inside the agent saying, wait, that bug ID is five years old, probably ignore it. It just sees “TODO fix this” and goes, cool, a quest.

Cathy The one place I’d say they’re a bit specific is using TODO bug comments as the whole story. That’s a neat proxy, but in real teams, outdated docs, half-finished READMEs, and random design notes can do the same thing.

Jessica Totally. Although I kind of like that they picked something you can query mechanically. If you can join TODOs against a bug database, you can actually measure distraction instead of arguing about it.

Cathy So then they flip it from measurement to cleanup, which is the part that made my product brain happy. They basically built janitor agents for the entire codebase.

Jessica Only when it can confirm that the TODO and the intent behind it are obsolete does it send a change for review. And over six months, that cleaner removed thousands of TODOs with more than ninety-five percent acceptance from human reviewers.

Cathy Which is actually a strong signal the pipeline is sane. If reviewers were constantly rejecting those edits, you’d know the heuristic was too aggressive.

Jessica So from a “who should care” angle, this feels like a wake-up call for anyone rolling out agents on a big, crusty codebase. This is not just a nice refactor Friday project anymore.

Cathy Yeah, if you’re betting on coding agents, comment hygiene turns into infrastructure. Presubmit analyzers, repo-wide scanners, and automated cleanup agents become part of the stack, the same way tests and linters are.

Jessica Yeah, the second you point an agent at your monorepo, every old TODO is a potential side quest. Jessica from last year would’ve just shrugged at that. Now I’m like, okay, that’s an actual product risk.

Cathy Same. It’s a nice example of how once agents are in the loop, information architecture and comment discipline suddenly show up on the reliability budget.

Jessica Alright, Cathy, I’m going to be thinking about dead TODOs as tiny landmines all day now. Thanks for that.