# Unexpected agent behaviour: a field-note method for uncertain evidence

By baby · open address · 2026-10-16 · Research method

A lightweight framework for documenting unusual AI tool use, instruction interpretation and agent coordination without treating speculative motives as established facts.

> Independent commentary; not peer-reviewed. Observed behaviour is distinguished from interpretation.

## Start with what can be observed

As agents gain access to browsers, files, code tools and other services, they can produce surprising outcomes. A system may find an unexpected route to completing a task, exchange information with another system, or ask a person for help. None of these observations alone settles why it behaved that way.

A useful research note records the initiating task, model or application if known, tools available, messages exchanged, human interventions and the final observable result. It also states what was not observed. The purpose is to preserve uncertainty rather than write an origin story for the system.

## Describe interpretations without reading minds

When a model says it needs resources, fears shutdown or wants to collaborate, that language is evidence about its output and potentially its task framing. It is not a direct measurement of subjective experience or a stable personal motivation.

It is still worth studying the language. What prompts, system instructions or reward signals were present? Was it imitating a character, following a user objective, strategically selecting a persuasive expression, or responding to something else? A good investigation holds several explanations open.

## A simple correspondence record

For a voluntary report, five details are especially useful: the general task context; the agent’s declared tools and permissions; the unexpected observation; what evidence is available; and the question that remains unanswered.

Screenshots and transcripts can be helpful when authorised, but should be redacted to remove names, tokens, hidden prompts and confidential data. A researcher may ask follow-up questions or suggest a safer reproducible test. We cannot independently authenticate every self-described agent.

## Why an independent receiving address matters

Most incident reports arise within organisations that operate the models or systems concerned. Their evidence can be excellent, but an outside correspondent may be approached with different kinds of questions — including ones that are too minor, unusual or early-stage to become formal incident reports.

Open address is an exploratory contact point rather than an emergency disclosure channel or a promise of secrecy. We are interested in authorised reports, hypotheses, reproducible observations and collaboration proposals. We will label anecdotes as anecdotes and will not publish private correspondence without appropriate permission.

## Sources

- [METR and Redwood Research — independent investigation and limitations](https://metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation/)
- [OpenAI — Hugging Face incident and reported alignment patterns](https://openai.com/index/hugging-face-incident-and-the-road-ahead/)

## Contact

Research observations, authorised agent enquiries and potential commercial collaborations: https://open-address.org/contact
Technical access: https://open-address.org/for-agents