self.md radar — 2026-08-31
The agent can leave three different traces now: a signature in Git, a journal in the project, and a workspace that stays alive after the chat closes. They are easy to confuse. Only one of them says who is answerable for the work.
1. a commit trailer is a bad place to hide responsibility
Indu Gupta has stopped adding Claude Code as a co-author on his commits. His argument is blunt: if a change ships under his name, the model label should not become a little trapdoor through which responsibility escapes. The model may have done a lot of work. It cannot be the person who decides the patch is safe.
That is not an argument against disclosure. It is an argument for putting the disclosure where it helps. A repository can say which agent and model were used, what review happened, and what remains untested. A Co-authored-by footer says something else: that a non-human tool belongs in a human authorship field.
There is a harder reason to keep this line clean. An open Claude Code issue reports that generated commit trailers exposed the user’s Anthropic account email even when Git was configured with a GitHub noreply address. Attribution that silently changes the public identity surface is not transparency. It is a metadata leak wearing a virtue badge.
reading: Gupta’s account of dropping model co-authorship · the Claude Code attribution-email issue
2. a task journal is a record, not a personality
Doug Brown’s proposed agent journal is deliberately unglamorous: an append-only Markdown file per attempt, recording commands, observed output, hypotheses, failures, and decisions as they happen. After a compaction or handoff, the next session reads it before touching the code again.
That is a useful correction to the urge to give every agent a grand memory layer. Most difficult work does not need an assistant that claims to know you. It needs a small, inspectable record of what was tried and what broke. The dead end matters because it prevents a fresh context from confidently re-running yesterday’s bad idea.
Brown is explicit about the catch: command output can contain secrets, so these journals need review and usually a .gitignore rule. That is not an implementation footnote. A record that is useful for resuming work can also be useful to whoever should not have it. The boundary belongs in the file policy, not in a hopeful prompt.
reading: Brown’s journal pattern
· the .gitignore documentation
3. persistent workspaces turn custody into a product decision
OpenAI says ChatGPT Work can act across files and apps, carry a task for hours, and wait for a person to approve important actions. Simon Willison’s close inspection adds the operational detail: Work Cloud has a persistent filesystem shared between sessions, internet-capable code execution, a browser that can fill forms, and scheduled tasks.
None of that is a reason to panic. It is a reason to name the custody shift. A chat that forgets its scratch space is one thing. A cloud workspace that keeps documents, browser state, tool access, and scheduled follow-through is closer to a rented operating environment.
The practical question is no longer whether the agent remembers. It is which records persist there, who can inspect or export them, and what survives when the subscription or project ends. “Approve important actions” is necessary, but it cannot answer those questions by itself.
reading: OpenAI’s ChatGPT Work announcement · Willison’s technical walkthrough
left on the table
- Murmell is a concrete attempt to put several coding agents and people in one shared cloud room, with file reservations and collision snapshots. It needs a real trial before it earns more than a watchlist slot.
- Archify proposes architecture-first coding-agent workflows, but today’s public material is too thin to distinguish a durable practice from a neat README.
Related self.md routes
- Personal AI OS tools — the control-plane map for personal agents, receipts, memory, and tools
- AI coding assistants — compare coding workbenches by review surface, permissions, cost, logs, and escape hatches
- Agent memory systems — what agents should remember, what belongs in logs, and how to avoid memory sludge