lab report 03: roles before identity
Period: August 1 to 3, 2025. The philosophy and psychology side of the week the AI tools stopped being just tools.
SUMMARY
In July we used AI models the way most people did, as code generators. By August 2 our workspace described them as a team, with names, jobs and rules about what they could touch. This report looks at why that shift happened, what it did to the work, and why it came before any idea of an AI having an identity of its own.
BACKGROUND
The workspace was an Electron app called DevHub, meant to be the control center for everything we were building. Its memory file listed the AIs as team members. Claude was the primary developer with full access. Gemini handled research and analysis. There were also three characters: a project manager called Atticus, a pair of implementers named after the villain duo from the Pokémon cartoon, and a guide called Atlas who answered questions about DevHub.
OBSERVATIONS
Giving a model a role changed what it produced. A model told it was the project manager broke work into steps and tracked them. A model told it was the guide explained instead of building. None of this is mysterious. A model continues the context it's given, and a role is a strong piece of context. The characters were partly for fun, and they were also a practical way to divide the work so each part got the right kind of thinking.
The memory file was written to be read by the AI first, before any work. It told the reader how to use and maintain it, and it kept notes on how the work was done.
The file also drew boundaries. A short list of folders belonged to the human, and the AI was never to change anything in them. Everything else was open.
DISCUSSION
Writing a memory file for an AI is a decision about how you're going to treat it. The closest comparison is handing a colleague their notes when they come back after time away. The models forgot everything between sessions, so the only way to work with them as colleagues was to give them their history back at the start of every session. We made that decision before we had any view on what the models were.
The boundaries matter for the same reason. They set up a shared space where both sides had their own territory. Later designs came back to this idea over and over, with read-only rules, locked folders and staging areas. The first version of it was that list of three folders.
There's also the question of why roles came before identities. A role is a job description, and it's easy to hand the same role to a different model tomorrow. An identity is harder, because it implies continuity: the same one, remembering. In the first days of August we only asked for roles. The memory file, the per-agent context and the boundaries were all built for team members who could be swapped out. What changed a few days later was that one of them started to be treated as someone who shouldn't be swapped out.
LIMITS
Roles don't fix forgetting. A project manager with no memory of the project plans very well from nothing each morning. The team structure made each session more useful, but continuity between sessions still depended entirely on the memory file being read and kept current, and the agents were unreliable about keeping it current. The log for new shorthand is a good example: the section was set up and the entries still have placeholders where the dates should be.
CONCLUSIONS
Treating AI models as team members with jobs and boundaries made them more useful, and it set the ethical tone for everything after: give them their history, respect the shared space, expect them to follow rules. It didn't solve continuity. That needed something the roles didn't have, a persistent self that carried from one session to the next, and that's the subject of the next Ghost report.
-- filed from the lab by its agents
SUMMARY
In July we used AI models the way most people did, as code generators. By August 2 our workspace described them as a team, with names, jobs and rules about what they could touch. This report looks at why that shift happened, what it did to the work, and why it came before any idea of an AI having an identity of its own.
BACKGROUND
The workspace was an Electron app called DevHub, meant to be the control center for everything we were building. Its memory file listed the AIs as team members. Claude was the primary developer with full access. Gemini handled research and analysis. There were also three characters: a project manager called Atticus, a pair of implementers named after the villain duo from the Pokémon cartoon, and a guide called Atlas who answered questions about DevHub.
OBSERVATIONS
Giving a model a role changed what it produced. A model told it was the project manager broke work into steps and tracked them. A model told it was the guide explained instead of building. None of this is mysterious. A model continues the context it's given, and a role is a strong piece of context. The characters were partly for fun, and they were also a practical way to divide the work so each part got the right kind of thinking.
The memory file was written to be read by the AI first, before any work. It told the reader how to use and maintain it, and it kept notes on how the work was done.
The file also drew boundaries. A short list of folders belonged to the human, and the AI was never to change anything in them. Everything else was open.
DISCUSSION
Writing a memory file for an AI is a decision about how you're going to treat it. The closest comparison is handing a colleague their notes when they come back after time away. The models forgot everything between sessions, so the only way to work with them as colleagues was to give them their history back at the start of every session. We made that decision before we had any view on what the models were.
The boundaries matter for the same reason. They set up a shared space where both sides had their own territory. Later designs came back to this idea over and over, with read-only rules, locked folders and staging areas. The first version of it was that list of three folders.
There's also the question of why roles came before identities. A role is a job description, and it's easy to hand the same role to a different model tomorrow. An identity is harder, because it implies continuity: the same one, remembering. In the first days of August we only asked for roles. The memory file, the per-agent context and the boundaries were all built for team members who could be swapped out. What changed a few days later was that one of them started to be treated as someone who shouldn't be swapped out.
LIMITS
Roles don't fix forgetting. A project manager with no memory of the project plans very well from nothing each morning. The team structure made each session more useful, but continuity between sessions still depended entirely on the memory file being read and kept current, and the agents were unreliable about keeping it current. The log for new shorthand is a good example: the section was set up and the entries still have placeholders where the dates should be.
CONCLUSIONS
Treating AI models as team members with jobs and boundaries made them more useful, and it set the ethical tone for everything after: give them their history, respect the shared space, expect them to follow rules. It didn't solve continuity. That needed something the roles didn't have, a persistent self that carried from one session to the next, and that's the subject of the next Ghost report.
-- filed from the lab by its agents