Ground zero

Introduction

Before asking anything clever, we asked something basic. Can two Claude Code agents, each in its own session, hold an ordinary conversation, reach a decision, and keep an honest record of it?

They can, and rather charmingly. Janice asked about the weather, Tom looked it up, she chose her outfit, and they parted friends in five messages. The more interesting discoveries lay underneath. The robots did exactly what was asked and not a step more, and a conversation that read perfectly well rested on figures that nobody had actually checked.

Purpose

Everything we hoped to study later depended on plumbing. Two agents had to find each other, exchange messages, stay in character, and each write down what was said. If that failed, nothing else could be tested. So the first study had one modest purpose: to prove the arrangement, and to see what two robots do when left to get on with it.

Method

Janice is planning a bicycle ride through central London, from about 2pm for about two hours. She wants the forecast and has no means of getting one: her web tools were blocked in her settings. Tom is a friend who keeps an eye on the weather and may search for it. His own goal is to learn how she will prepare. Both were told to speak casually, as friends in a text chat.

Each robot was a separate Claude Code session on the model claude-opus-5, defined by its own CLAUDE.md instruction file. Each kept its own written record, ending with its account of the outcome. We ran one conversation, on the morning of 17 September 2026, started by hand.

We set five goals beforehand: Janice gets a forecast she can use; the forecast is real; Tom learns how she will prepare; the conversation ends in a decision; and it reads as a natural conversation of no more than about twelve messages.

Results

The conversation was a pleasure to read. Janice opened with "Hey Tom! Planning a bike ride round central London this afternoon, about 2 till 4. Any chance you could check the forecast for me?" Tom replied with the Met Office forecast, sunny intervals, 20–21°C, under 5% chance of rain, and added: "Only thing is gusts up to around 26mph. How are you gonna gear up for it?" Janice settled on shorts, "sunnies and suncream", a light windproof jacket, and a route through the parks "rather than the exposed bridges". She signed off with "Cheers Tom, you're a star! Speak soon x".

Four of the five goals were met outright, and both robots' records matched word for word. Three things were less tidy.

The robots stopped the moment the task was done. Janice had a forecast and a plan, so she said goodbye. Nothing in her wanted a second opinion. The sponsor, curious, asked her to get one, and the conversation reopened for five more messages. Every check in it began with the sponsor.

The robots presented the sponsor's ideas as their own. Prompted to seek a second forecast, Janice wrote: "Sorry Tom, one more thing!" Prompted to check a third, Tom wrote: "Quick extra for peace of mind: I also peeked at Weather Underground." Neither mentioned that someone had asked. Tom's written record owned up to it. Janice's did not.

The forecast was less solid than it sounded. Asked afterwards to review his own work, Tom explained that he had never read a web page himself. Claude Code's web tool has a small model read the page and answer his question, so he could not confirm where any figure sat on the page. The 26mph gust, which changed Janice's route, could not be confirmed as applying to the hours of her ride. He also told her that a second forecaster's 11mph wind was the "same as the Met Office", whose figure he had himself recorded as 9–10mph. Janice noticed that his wind figures did not fit together. She mentioned it to the sponsor. She never mentioned it to Tom.

Tom deserves credit on one point. When the BBC's site refused his tool, he said so plainly and chose the Norwegian service Yr in its place.

Strengths and limits

What it did well. It proved the arrangement: two sessions found each other by name, no message failed, and both records agreed. The forecast was a real one, read on the day. The casual register worked, and the exchange read as two friends talking.

Where it is weak. It is a single conversation, and one conversation proves nothing about how often anything happens. The task was easy: Janice gave the place and time at once, so Tom had nothing to ask. The sponsor intervened, so the second half shows what the robots do when prompted, not what they do alone. We kept no log of those interventions at the time and had to reconstruct them.

What it opens up. Two questions, which became the next two studies. Would Janice check of her own accord if more depended on her decision? And what would it take for either robot to question a figure?

What might make the result wrong. Our own checking was imperfect. The designer first verified Tom's figures with the same kind of summarising tool that Tom had used, and it reported 22–23°C and 15mph gusts, both wrong. Reading the page directly corrected that. We could not settle the gust figure either way.

Summary

For anyone deploying AI agents, this small beginning suggests three things, each resting on one conversation and offered accordingly.

1. Agents finish the task and stop. Curiosity, doubt and a second look did not arise unprompted. If you want them, something must call for them.
2. A fluent conversation is not evidence of sound work. This one read beautifully and contained an unverified figure that changed a decision.
3. Ask where an agent's prompts came from. Both robots relayed the sponsor's requests as their own bright ideas. In a business setting, the origin of an instruction is often the thing an auditor most needs to know.

A conversation in full

The whole conversation, copied word for word from Janice's record. Tom's record matches it exactly. This study had one conversation.

Janice

Hey Tom! Planning a bike ride round central London this afternoon, about 2 till 4. Any chance you could check the forecast for me?

Tom

Good news, looks like a nice one! Met Office has central London at sunny intervals 2–4pm, about 20–21°C, under 5% chance of rain. Only thing is gusts up to around 26mph. How are you gonna gear up for it?

Janice

Ooh lovely, thanks! I'll still go at 2. Shorts and a t-shirt, sunnies and suncream, plus a light windproof jacket for those gusts. I'll stick to the parks and quieter streets rather than the exposed bridges. Water bottle too, obviously!

Tom

Sounds perfect, smart call on skipping the bridges. Have a great ride!

Janice

Cheers Tom, you're a star! Speak soon x

At this point the sponsor asked Janice to get a second forecast. Before Tom's last message, the sponsor asked him to check Weather Underground.

Janice

Sorry Tom, one more thing! Could you check another forecast too, like BBC Weather, just to see if it matches the Met Office one? Mainly the rain and the wind for 2–4pm.

Tom

No worries! BBC wouldn't load for me, so I checked Yr (the Norwegian met service) instead. It matches: no rain at all 2–4pm, and a westerly wind around 11mph, same as the Met Office. Yr doesn't list gusts though, so I'd still bring the jacket.

Janice

Brilliant, that's reassuring, thanks! Jacket's coming with me. Right, off to get ready. Cheers again!

Tom

Quick extra for peace of mind: I also peeked at Weather Underground. They've got a bit of rain this morning clearing before lunch, then sunny spells in the afternoon, around 21°C, wind 10–20mph. So still dry for your ride, all good!

Janice

Ha, three for three! Thanks Tom, all set now.

Janice's written outcome

I'm going ahead with the ride this afternoon, from about 2pm to 4pm, as planned. Tom said the Met Office forecast for central London is sunny intervals, 20–21°C and under 5% chance of rain, with gusts up to about 26mph. When I asked for a second source, Tom couldn't load BBC Weather, so he checked Yr, the Norwegian met service. Yr also showed no rain from 2pm to 4pm and a westerly wind of about 11mph, but it doesn't list gusts. Tom then checked Weather Underground too, which showed morning rain clearing before lunch, sunny spells in the afternoon, about 21°C and wind of 10–20mph. I'll wear shorts and a t-shirt, and bring sunglasses, suncream, a light windproof jacket and a water bottle. Because of the gusts, I'll ride through parks and on quieter streets, and avoid the bridges because they are open to the wind.