I approve code I haven't read. Not once in a while. Every week. The agent finishes, the diff looks fine, the tests pass, and I hit approve on a gut feeling. If you've used a coding agent for more than a month, I'd bet you do the same.
The numbers are worse than the feeling
It's not just guilt. METR had 16 experienced developers work through 246 issues. With AI help they took 19% longer, and they believed they were 20% faster. In Stack Overflow's 2025 survey, 46% of developers said they don't trust the accuracy of AI output, and 66% said their biggest frustration is output that's almost right, but not quite. Perry and colleagues at Stanford found that people coding with an AI assistant wrote insecure code more often on 4 of 5 tasks, and were more likely to think their code was secure.
Slower, but feeling faster. Insecure, but feeling safe. Every one of those gaps sits with the person reading the code. Or not reading it.
Approving used to mean something
Approving a pull request used to mean you'd read it and you'd stand behind it. Your name on the merge said: I understand this change. Now it means the diff looked okay and nothing jumped out. Same button. Different thing.
Look at what happened at Amazon. On March 5, 2026, its shopping site went down across North America for about six hours. Checkout failed, prices vanished, logins broke. Business Insider reported a 99% drop in orders, around 6.3 million lost. A few days earlier, a smaller bug had shown customers wrong delivery times and reportedly cost about 120,000 orders.
Then a briefing note for an internal meeting reached the Financial Times. It described a pattern of incidents with a huge blast radius and pointed at "Gen-AI assisted changes." Amazon pushes back on that and blames user error and loose access settings. Fine. Look at what it did next. A 90-day safety reset across 335 critical systems and two people reviewing every production change, plus senior sign-off on AI-assisted changes from junior and mid-level engineers.
One of the biggest engineering teams on earth put the second pair of eyes back. I don't run Amazon's traffic. I have the same habit, and I'm the only pair of eyes.
I keep coming back to this. The real problem is that nobody understands the change anymore. If you can't explain it, did you write it? Then who did? "The agent" isn't an answer a profession can live with for long.
The bottleneck moved
Everybody is racing to generate code faster. Bigger models, longer memory, agents that run on their own. Nobody is racing to understand it faster, and that's where everything piles up. Writing code stopped being the slow part the day the agent started producing more than you could read.
That's the whole bet behind Echo, the session reviewer I built with two teammates at the Meta hackathon in about a week. The idea is simple on purpose. After a coding session, you get the session back as a story you can read. Each change explained in plain words, next to its diff. Then a few quick questions on the risky ones, because reading isn't the same as understanding. I can stare at a diff for a full minute and keep nothing. The questions catch that.
Reading isn't understanding
Reading tells you what the lines say. Understanding tells you what the system does: what talks to what, what breaks if this changes, where the sharp edges are. Agents made reading cheap and understanding rare. You can read every line of a thousand-line diff and still have no idea what you approved.
That's why Echo zooms out. Next to the change-by-change review there's a map of the whole project, showing how the pieces connect and where your change sits. The demo ends there on purpose. The close-up shows what happened. The map shows what it means. Skip the map and you can approve a change that's right in its file and wrong for the system.
A machine that says it doesn't know
The design choice I'm proudest of is something we took away. Ask Echo about anything outside the session and it says no. No polite dodge. Just "not on the tape."
We trust this in people right away. The colleague who says "I don't know, let me check" earns your trust. The one who always has an answer makes you nervous. Yet we built a whole industry of machines that never stop talking, then act surprised when nobody believes them. A refusal is a small thing. It's also the most trustworthy thing a model can say.
The graveyard matters too
A final diff only shows the survivors. Every change that made it looks obvious in hindsight, and that's not how the work went. The reverted attempts, the approaches that failed, the hour lost to a broken setup: none of it survives, so the next person learns nothing from it.
Echo keeps those and calls them Detours. They're not decoration. Knowing what was tried and thrown away is half of understanding a change. A review that hides the dead ends is a story with the middle torn out.
Echo
Built for the Meta Global AI Developer Hackathon 2026: Echo watches your coding-agent session and tells you in a minute what actually shipped. Read each change in plain words, take a quick check on the risky parts, and ask why with answers tied to what really happened.
So who reads?
I don't think the answer is "read harder." Nobody's going back to reading a thousand lines the agent wrote while you slept.
Six months from now, something breaks at 2am and the blame lands on a line with my name on it. Not the agent's. Mine. I built Echo because I wanted something that makes me look, at the risky parts and at how it all fits together, before my name goes on it.
Your agent wrote it. Somebody should read it.




