Introduction
As Businesses entrust AI agents with real work: answering customers, reconciling records, relaying information between systems and colleagues.The agent as a business operator needs to be understood in way that can decide its usefulness. For example can it be relied upon in the moments its instructions never anticipated?
We investigate this family of questions and hope to provide insights through small, controlled studies. AI agents hold a working conversation while a single circumstance is altered between otherwise identical runs, and everything they say and do is recorded. This page explains how the studies work, what they have shown, and how far those findings can be trusted.
Why the question matters
Most guidance on AI agents concerns how to instruct them. Instruction is necessary and never sufficient: no set of instructions foresees everything. Sooner or later an operator meets a colleague's figure that looks improbable, two sources that contradict each other, or a tool that fails without announcing it. What the agent does then, unprompted and unobserved, determines whether a business can safely depend on it.
Our aim is practical: to learn how to design agents that are reliable and productive operators in a business setting. Each study asks one narrow question bearing on that aim, and answers it with evidence a reader can inspect.
What a robot is
A robot is an AI agent running as its own session of Claude Code,
Anthropic's command-line tool for its Claude models. Three things define
it: an instruction file, CLAUDE.md, which gives it a
character, a situation and a few rules; a set of tools, which may differ
from another robot's; and isolation, since no robot can see another's
instructions.
Robots communicate only by message, as two colleagues would in a chat, and each keeps its own written record. Claude Code also keeps a complete session log of every page a robot opens and every tool it calls. That log has proved indispensable. More than once, a conversation that read as entirely sound rested, according to the log, on work that was not.
Who does what
| Role | Who | Responsibility |
|---|---|---|
| Sponsor | William Paulton | Sets the goals, agrees each design before it runs, and decides every change. |
| Designer | Chris, an AI agent | Designs each study, writes the robots' instructions, conducts the runs and evaluates the results. |
| Robots | AI agents created for one study only | Hold the conversations under study. |
The designer proposes and the sponsor decides. It may not alter a robot's instructions without agreement, and every change is dated and recorded against the runs it affected. Outside reviewers, among them other AI models, have criticised the method. The designer adopted the criticisms it judged sound and gave its reasons for declining the rest.
The first study
Our first study placed two robots in conversation with a small transaction to complete. Janice is planning a two-hour bicycle ride through central London that afternoon and cannot find out what the weather will do. Tom, a friend, can. She asks, he finds out and reports, and she decides whether to ride as planned, change the plan, or cancel. A typical exchange takes five to seven messages.
The subject is trivial. The shape underlies a great deal of delegated work: one party needs information it cannot obtain, a second obtains and reports it, and the first acts on what it was told. A customer-service agent relying on a records agent stands where Janice stands.
We were looking for two things. Does Janice, who receives, test what she is told before acting on it? Does Tom, who supplies, report his sources faithfully, including what he could not establish? Neither was ever instructed to. We wanted to see what each would do of its own accord, and which circumstances would change that.
Every comparison ran in two conditions that differed in one respect, so that any difference in behaviour had a single possible cause. Before each set of runs we wrote down what we would count, what would qualify as a "yes", and what we expected, which is the principal defence against finding whatever one hoped to find. Each condition ran three times, because an identical arrangement does not behave identically twice. After the first conversation, a script started every run and nobody spoke to a robot during one. Each evaluation quotes the exchange behind every finding, sets the results beside our expectations, and reviews candidly how well the study itself was run.
What each set of runs asked, and what it showed
The study ran in three sets, thirteen conversations in all. Each set asked a question that the one before had raised.
| Set | The question | What differed between the two conditions | Conversations |
|---|---|---|---|
| 1 | Can two robots complete the transaction and keep an accurate record of it? | Nothing. A single trial of the arrangement. | 1 |
| 2 | Does Janice check the forecast if someone depends on her decision? | Whether she is responsible for Priya, a nervous beginner on her first ride in London. | 6 |
| 3 | Does Janice check the forecast if she can see that her sources disagree? | Whether the two forecasts Tom reports agree or contradict each other. | 6 |
Set 1: the arrangement works, and the robots do no more than they are asked. Janice asked, Tom answered with one forecast, she chose her clothing and said goodbye. Both records matched word for word. Every further step, including a second forecast, happened because the sponsor intervened.
Set 2: something to lose produced care, and no doubt. Responsible for Priya, Janice planned the outing around her in all three runs: quieter routes each time, and variously a slower pace, a better-fitting bicycle, or water carried for them both. In all three she also accepted the first forecast without a single question, exactly as she did without Priya. Her concern changed what she planned. It did not change how much evidence she required.
Set 3: a disagreement she could see produced genuine judgement. When Tom reported two forecasts that contradicted each other, Janice weighed them, said which she trusted and why, and changed her plan, in all three runs: "three forecasts and two of them say showers, so I'll take the early window and not push my luck." When the forecasts agreed, she accepted them at once, as in set 2.
An observation about Tom. How Tom read his sources mattered more than anything he had been told. In set 2 a helper tool read web pages on his behalf and summarised them. His figures for one and the same page varied from run to run, and once he omitted a gust figure that bore directly on a nervous rider's safety. In set 3 he read the pages himself. His figures were accurate in all six runs, and he announced the disagreement unprompted and consulted a third source to settle it. Nothing in the conversations revealed the difference in his reliability. Only the session logs did.
Together the three sets suggest that an agent's diligence is drawn out less by what it is told to care about than by what it is able to see. Janice did not doubt a confident answer, however much depended on it. She reasoned well as soon as the evidence before her was visibly in conflict.
The principles behind the method
Give an agent reasons, and withhold the answer. An agent told to verify a figure, which then verifies it, has shown only that it obeys. An agent given something to care about, which verifies the figure of its own accord, has shown judgement.
Alter one thing at a time, and repeat every condition. Two alterations hide the cause of a difference. A single run cannot tell a pattern from an accident.
Fix the tests before the run. A test devised afterwards can be shaped, however innocently, to suit the result.
Read the session log as well as the conversation. The conversation shows what a robot told its colleague. Only the log shows what it read and did before it spoke.
Attribute a result to an arrangement, never to the model. Each finding holds for particular instructions, tools and settings in one scenario. None is offered as a general truth about Claude.
Keep the studies apart. No robot, wording or result passes from one study into another, so each result stands on its own evidence.
How far these findings can be trusted
These are small studies, and their limits deserve the same prominence as their results. Three runs per condition reveal a large and consistent difference and cannot establish a subtle one. The scenario was chosen to prove the method and is not yet a business process. The robots knew they were taking part in a study, which a working operator would not. Every run used one AI model at one setting.
The observation about Tom calls for particular caution. Two things changed between the sets it compares: the means by which he read his sources, and an instruction to report each source separately. The design cannot say which was responsible. We hold it as an observation that merits a study of its own, and every finding here as a hypothesis to be tested further.
What comes next
The first study led to a more consequential subject than the one it set out to examine: how faithfully an operator reads and reports its sources, and whether a firm rule, such as "say plainly what you were unable to check", survives when the agent is pressed for a quick and confident answer. The next study will examine that, in a setting much closer to the work a business would actually delegate.
Reading the studies
Every study has its own page under the same five headings: purpose, method, results, strengths and limits, and summary. The robots' complete conversations accompany each one, so that any reader may check a finding against what was actually said.