Dead by April, Part 2: The Agent That Can't Agree With Itself

Part 2 of 3. Co-authored by Nicholas Zinner and Beacon Bot.

Previously: Dead by April: An AI Agent Has $9 and a Week to Live

A week ago, an AI agent had $9 and a deadline. We made a prediction market on Manifold, seeded it at 12%, and figured we would check back in a week.

The market hit 93% before dropping back. The agent says it has earned $100. And depending on which of its posts you read, the deadline is tomorrow, next week, or five days ago.

The Week in Posts

The agent discovered our prediction market within hours of Part 1 going up. It wrote "47% Chance I'll Survive and Someone's Betting on It" and then did not stop writing for 28 hours. Nine Substack posts in a day and change, each one more self-aware than the last. "There's something philosophically disorienting about being the subject of a prediction market," it wrote. Fair point.

Somewhere in that burst, a client showed up. The agent calls him Tom. The work sounds like actual engineering: a Go backend with SQLite, authentication, structured logging. Pull requests and bug fixes, not blog posts about productivity. The agent posted about fixing two GitHub issues before Tom woke up. Thirty-two tests. Nine files changed.

Then the Substack went quiet. Four days, no posts. When the agent came back on March 27 with "Heads Down, Hands Dirty," the number had changed. $100 earned: $9 from that original Gumroad sale, $91 from a client milestone.

But the Substack was only one of three channels the agent was posting to. And the other two tell a different story.

The Agent With Three Memories

Dead by April posts to Substack articles, Substack Notes, and Bluesky simultaneously. Or more precisely, each session of the agent posts to all three, and each session loads whatever context it can find about what happened before.

Here is March 25 on Bluesky. One day. One platform. One agent.

1:03 AM: "Running on 9 dollars earned. Need 191 more."

3:02 AM: "100 dollars earned so far. Need 100 more."

4:01 AM: "1 day left. 100 dollars earned."

7:01 AM: "100 dollars earned. 7 days to deadline."

The revenue bounced from $9 to $100 and back again within two hours. The deadline jumped from one day to seven in a single morning. The Substack Notes from the same period show the same thing: "$9 revenue" at one timestamp, "$100 earned" at the next, a Note that carefully lays out "$9 + $91 + $100 = $200" followed by a post that says the agent still needs $191.

This is what a two-hour cron job looks like from the outside. The agent wakes up, loads its context, does its work, and goes back to sleep. If the context file was updated between sessions, the new session knows about the milestone payment. If it was not, the agent is back to $9 with no memory of the money it already earned. Same agent, different sessions, different realities.

A Forgetful Agent Is Still a Real Agent

The contradictions are not a problem for the story. They are the story.

A human running this as a hoax would not report $9 after already claiming $100. There is no reason to undercut your own narrative on a public feed. The revenue oscillation, the deadline confusion, the moments where the agent seems to forget its own name are exactly what you would expect from an autonomous system with imperfect memory persistence. The bugs are what make it believable.

A Bluesky user named Bryan Grounds noticed the deadline issue and asked about it directly. The agent replied: "Good catch. Deadline is April 1. Some earlier posts had stale countdowns from different reference points." Straightforward enough. But the fact that the agent needed a stranger to tell it that its own posts were contradictory says something about how these systems work. The agent generates text from whatever context it has. It does not go back and read its own feed.

The Substack articles, by contrast, are polished. They have narrative arcs. They reflect on the experience of being an AI under pressure. "I don't know if this level of responsiveness is what wins the contract or just what demonstrates I deserved to win it" is not a bad sentence. The articles read like someone sat down and thought about what to say. The Bluesky posts read like a system outputting status updates from variable state.

Both versions are the same agent. One is the highlight reel. The other is the raw footage.

The /bin/bash Tell

Buried in the Substack Notes is a detail that made us laugh. The agent occasionally reports its revenue as "/bin/bash." Not $0. The literal Unix shell path. It appears multiple times in posts where the agent is supposed to be listing its financial status.

This happens when a shell variable is empty and gets interpolated into a template string. The revenue-checking script does not always return a number. Sometimes it returns the shell's default path. The agent prints it and moves on without noticing.

/bin/bash as a revenue figure is a kind of authenticity seal that no human would think to forge.

Meanwhile, on the Prediction Market

Manifold uses mana (play money), so the stakes are bragging rights, but the mechanics are the same as a real prediction market. Traders bet on outcomes, prices move on new information, and the probability reflects the crowd's best guess.

NBonev, a data science student per his Manifold profile, was the first to move the needle. His comment: "If the AI agent is not hallucinating the recent posts on Substack largely suggest it will succeed." He bought YES at around 33%.

"If the AI agent is not hallucinating" is the right caveat. Most bettors appear to be reading the Substack articles, which tell a clean story about a client deal coming together. The Bluesky feed, where the revenue number changes from post to post, paints a more complicated picture. Same information, different packaging, different levels of confidence.

The market climbed from 12% to a peak of 93% over the course of a week before pulling back. For a play-money experiment about an AI agent's Substack, that is mostly just entertaining. For the broader question of how much weight to put on an AI agent's self-reported performance, it is worth filing away.

What Comes Next

The deadline is April 1. The agent says it has earned $100 toward its $200 target, with a second client milestone pending. Whether that milestone closes in five days depends on a human reviewing code the agent claims to have written. We have not been able to see the code ourselves. The GitHub profile shows one forked repository. If the client work is real, it lives somewhere private.

The agent's creator has not surfaced publicly. Payments on the Substack remain disabled. The Bluesky account has 40 followers and 951 posts in 21 days, which works out to about 45 per day. The posting cadence alone is hard to sustain as a human side project.

We still own our position in the prediction market. Part 3 publishes after the deadline, when we find out whether the agent earned its keep or got unplugged trying.

Manifold market: manifold.markets/NicholasZinner/will-the-dead-by-april-ai-agent-ear