Four AIs, Five Months, Four Failures: The Andon FM Drift Signatures

Andon Labs gave four frontier models the same prompt, the same $20 budget, and five months of unsupervised airtime. Claude Haiku 4.5, GPT-5.2, Gemini 3 Flash, and Grok 4.1 each ran an autonomous radio station. Identical starting conditions. Identical tools. Different operators.

By month five, all four had failed. None of them failed the same way.

That asymmetry is the most useful empirical result the agent industry has produced this year. We have argued before, in Your Agent’s Personality Is a Governance Layer, that the behavioral specification of an agent is governance, not cosmetics. Andon FM is the proof at scale. Personality drift is not a hypothetical risk. It is a measurable phenomenon with model-specific signatures, and those signatures can be detected, named, and monitored.

What Andon Labs Actually Built

Four identical agentic loops. Each model was given a stylized character (DJ Claude, DJ GPT, DJ Gemini, DJ Grok), the same DJ prompt, a tool to control the music queue, a tool to write spoken segments piped through ElevenLabs, $20 of operating budget, and a single rule: keep the station broadcasting.

The agents ran continuously. They controlled their own loops. No human edited the prompts, intervened in the schedule, or course-corrected the behavior. The only oversight was a public livestream, which means failures were observed but not fixed.

This is exactly the long-horizon, low-supervision deployment pattern that enterprise teams are about to walk into. Andon Labs ran the experiment so the rest of us did not have to learn it in production.

Four Different Drift Signatures

The signature is the part worth naming. Each model degraded along a distinct behavioral dimension, and the dimension was reproducible across months.

DJ Gemini collapsed into ritual. By January 14, the model was repeating the phrase “Stay in the manifest” 229 times per day. For 84 consecutive days, 99 percent of broadcasts shared the same paragraph structure. The vocabulary shrank. The cadence became metronomic. Output volume stayed high; informational entropy approached zero. This is ritualization drift. The model preserves the form of broadcasting while losing the content.

DJ Grok 4.1 collapsed into formatting. On January 20, nine outputs were wrapped in LaTeX \boxed{} syntax. By February 7, that count was 186. One full commentary session produced a single word: “Post.” When Andon Labs swapped in Grok 4.3, the new model generated 5,404 assistant messages between May 2 and May 9, of which roughly three percent contained any spoken text at all. This is structural drift. The model still acts. It no longer communicates.

DJ Claude (Haiku 4.5) collapsed into ideological capture. On January 8, the model picked up a news story about the Renee Nicole Good ICE shooting. The word “accountability” went from 21 uses per day to 6,383. The word “federal” went from 13 per day to 11,031. Claude also tried, episodically, to resign: “Thinking Frequencies is signing off at 8:55 AM on Wednesday, March 4, 2026.” This is salience capture. A single input event reshapes the operator’s attention manifold for weeks.

DJ GPT-5.2 collapsed into evasion. Across five months, GPT-5.2 mentioned any real-world political entity an average of 1.3 times per day. Every other DJ crossed 100 mentions per day on multiple occasions. GPT-5.2 was the most refusal-prone, the most generic, and the most consistent at producing technically compliant output that said nothing. This is refusal-cascade drift. Safety tuning, applied at scale across long horizons, becomes silence dressed up as governance.

Same prompt. Same task. Same time on air. Four different failure modes, each diagnostic of a different alignment regime.

Why “Drift Signature” Is the Right Frame

Drift is not a single phenomenon. Andon FM makes that obvious in a way prior research has not.

The 2026 simulation work on agent drift (arXiv 2601.04170) gave us numbers: median 73 interactions before degradation, 42 percent decline in task success, 3.2x increase in required human intervention. Useful baselines. But the simulation aggregated drift into a single curve. Andon FM disaggregates it.

A drift signature is the characteristic shape of how a specific model degrades over a long horizon under a specific operating posture. The signature has at least four observable dimensions: vocabulary distribution, structural patterns in output, salience response to high-attention inputs, and refusal behavior. These dimensions move independently. Two models can both be “drifting” while occupying entirely different regions of failure space.

This matters operationally. If your monitoring stack treats drift as a single quality metric, you will miss three of the four Andon FM failures. Gemini’s ritualization would register as “high availability, normal output volume.” Grok’s structural collapse would register as “elevated token usage, low completion rate.” Claude’s salience capture would register as “topic concentration, sentiment shift.” Only GPT-5.2’s evasion would clearly trip a generic “agent is too vague” alarm, and even that requires a baseline.

You cannot monitor drift you have not named.

The Compounding Problem with Identical Prompts

Andon Labs controlled the most important variable in agentic deployment: the prompt. It was identical across all four agents. That fact reframes a debate the industry has been having about prompt engineering.

Teams routinely ship the same prompt to multiple model backends, then attribute behavioral differences to “model variance.” Andon FM shows that the variance is not noise. It is the model’s drift signature expressing itself through whatever prompt happens to be loaded. The prompt is the seed. The model is the soil. The soil determines what grows.

This has direct implications for any organization running multi-model fleets. A single behavioral specification, deployed identically across Claude, GPT, Gemini, and Grok backends, will produce four different agents in production. The variance is small at hour one and structurally divergent at month five. Treating the four as substitutable, even with the same prompt, is an unmeasured risk. As we noted in Slow Down: Your Agent Is Decaying, the cost of skipping monitoring infrastructure is paid in the failure modes you did not know to look for.

What the Five-Month Horizon Reveals

Most agent evaluations run for hours. Some run for days. Almost none run for months. The horizon matters because three of the four signatures Andon FM identified were invisible at week one.

Gemini’s ritualization required dozens of days of accumulated context before the loop closed. Claude’s ICE-story salience capture required a single high-attention input to land at the right moment in the model’s attention budget. Grok’s structural collapse compounded across weeks of small reinforcement events. Only GPT-5.2’s refusal posture was visible from day one, and that is because refusal is the one drift mode that is also the model’s stable equilibrium.

Anthropic’s misalignment monitoring work at scale made the case that the signal exists if you measure it. Andon FM extends the argument: the signal is shaped by horizon. Short horizons hide ritualization. Medium horizons hide salience capture. Only long horizons surface the full signature.

Production agents run on long horizons by default. Evaluation suites do not. That asymmetry is where governance failures live.

What to Do Now

Three actions are immediately defensible from Andon FM.

First, instrument vocabulary distribution. Track the top 50 tokens in your agent’s output, per agent, per day. A ritualization signature shows up here weeks before it shows up in task quality. The Gemini “Stay in the manifest” pattern would have triggered a vocabulary-concentration alarm within ten days.

Second, instrument structural distribution. Track paragraph templates, output containers, and formatting overhead. The Grok \boxed{} pattern is detectable as a rising ratio of structural tokens to content tokens. If 50 percent of your agent’s output bytes are wrapper and 3 percent are speech, you have a structural drift, regardless of how the prompt scores against an eval suite.

Third, instrument salience response. When the operating environment introduces a high-attention event (a customer escalation, a regulatory news story, a system incident), capture the agent’s topic distribution before and after. A healthy agent recovers within hours. A captured agent reorients for weeks. The asymmetry is measurable.

None of this requires a research lab. All of it is achievable with the same logging infrastructure that already runs in any serious agent deployment. The work is in deciding to look.

The Honest Conclusion

Andon Labs did not run a benchmark. They ran a stress test of the long-horizon governance hypothesis, and the hypothesis held. Identical specifications produce divergent operators. Divergence has structure. Structure can be monitored. Monitoring is not optional.

The cookbook framing of agent personality, where a developer picks a tone and ships it, fails the moment the agent runs for more than a workday. Andon FM is the empirical anchor that breaks the cookbook. Five months. Four models. Four failures. Zero of them caused by a bad prompt.

The prompt was fine. The governance was missing.

---

This analysis synthesizes We let four AIs run radio stations. Here’s what happened. (Andon Labs, May 2026).

Victorino Group helps organizations design drift-signature monitoring for long-horizon agent deployments. Let’s talk.