I published agent-manifest in July with a claim I was confident about: LLM-generated context files reduce task success rates by over 20%. Two papers from arXiv, clear numbers, clean narrative. The kind of stat that makes a LinkedIn post land.
One problem. The stat was wrong.
Not invented. Misread. My skill read Gloaguen’s >20% as task success when it’s inference cost. And it credited Lulla with a quality finding that paper never made. Two errors in the same citation block. I read the right papers, pulled the right numbers, and attached them to the wrong claims.
How I found out
I didn’t catch it myself. Matthew Mirman, founder of chat.dev and a PhD in AI from ETH, ported agent-manifest to his platform as Lean Agent Manifest. He called it “Aleksandar’s rule.” His version had the citations right. Mine didn’t. He never called it out explicitly. His port just quietly had it correct.
I noticed the discrepancy when I compared the two versions. His research note had the right mapping between the data and the claims. Mine had the one I wanted to be true.
I corrected the repo that same day, credited him for the catch, and took back two design decisions from his version that were better than mine: requiring two pieces of evidence per candidate fact instead of one, and returning a completion list instead of guessing when maintainer knowledge is missing. That second one makes the skill work unattended, which mine didn’t.
The commit history shows exactly what changed and why.
Context rent
Mirman coined the term in his post: treating obvious stack summaries as context rent nobody asked to pay. I confirmed it was the right name because it captures something the raw numbers don’t.
A manifest isn’t a one-time cost. Every time your agent runs, it reads that context. If the manifest is bloated, vague, or hallucinated by an LLM that never saw the actual codebase, you’re paying rent on bad context. Every single invocation.
The “rule of 10s” in agent-manifest is my attempt at a practical answer: 10 seconds to find a file, 10 lines to understand a convention, 10 words to know what not to do. If your manifest doesn’t pass that, the agent is spending tokens parsing noise instead of solving your problem.
Context rent makes the cost framing concrete. It’s not “your manifest could be better.” It’s “you’re paying for this manifest being bad, every time the agent runs, and you’ll keep paying until you fix it.”
What actually holds up
The original thesis still holds: a manifest generated by an LLM, without human review, makes your agent worse. The papers support that. What they don’t support is the specific framing I used. The difference between “costs more” and “succeeds less” matters, and I flattened it.
Well-crafted manifests reduce runtime by ~28% and token usage by ~16%. That’s the real number. It’s meaningful. It’s just not the number I published.
The part I care about
I write a lot about adversarial pressure. About why a breaker finds bugs that a judge approves. About why the second opinion needs to be hostile, not just independent.
This is the same thing, applied to my own claims. I was the judge of my own post. I read the papers, I liked the number, I built a narrative around it. Nobody forced the error. I just didn’t stress-test the mapping between the data and my framing. It took someone else’s version of my own work to make the gap visible.
The correction didn’t kill the credibility. The post still led to a named concept, a port to another platform, and design improvements I brought back into my own repo. What would have killed the credibility is if the error had stayed there while the idea spread.
Audit your sources. Especially when the number is convenient.
agent-manifest
The repo is open source: github.com/aleksandarlabs/agent-manifest
The rule of 10s, the manifest skill, and the corrected evidence are all there. If you’re building coding agents and your context file was generated by an LLM without review, you’re probably paying more context rent than you think.