Topics discussed:

• Nikolai proposed agentic workspaces: today you sit alone in a terminal with an agent, but real work happens in teams. The sketch is a virtualised, shared environment — rooms with at least one agent, a running Linux box, a skills directory — described declaratively the way you'd describe a Kubernetes workload, then cloned or rolled back like a commit. Part of the motivation is that non-programmers are arriving in numbers and drown in environment setup; the current answer is 'share your screen and I'll tell you what to tell the agent', which doesn't scale.
• The obvious objection came straight away: Zed and Cursor already have collaboration features. The answer was that those sessions live on one developer's laptop and die when they close it, which is fine for programmers and useless for everyone else — this is meant to be genuinely shared and to outlive any one participant. The harder problem he flagged is auth: once two people share an environment and one of them wants the agent to read mail or documents, you need a permissions layer that can grant access once, or for a window, or for specific actions — much trickier than tools installed locally for a single developer.
• A demo of an AI assistant inside a questionnaire builder, where tool calls are bridged to the application's own state so the agent can do anything a user could through the UI, plus PDF-to-questionnaire import. It also exposes Web MCP and bridges it out to desktop agents, so a CLI agent can drive the same page after the page asks the user for consent. The argument for going out to the CLI rather than keeping a panel in the app: the built-in assistant has no context management, so when the context fills you have to clear it and start over, whereas a CLI agent handles that automatically and brings its other tools along.
• Two context lessons from that demo. First, the models already know FHIR well enough that nothing extra is added to the context — but they do invent variables that don't exist in the FHIRPath evaluation context, so there's a tool that validates an expression's referenced variables and types before it can be saved, giving the agent a deterministic loop to iterate against. Second, the tools are deliberately surgical: small atomic updates to individual properties rather than regenerating the whole questionnaire each time, which saves context and enables incremental edits — at the cost of many tools, and every tool eats context too. The suggested resolution from the floor: this is the shape that eventually gets packaged as a skill.
• Malta showed Open Brain, a pluggable memory server (MCP, Postgres, pgvector) that persists what agents learn across sessions and projects. The distinction that carries the design: refine memories is rule-driven and automatic — dedupe, clean up, agent does what it likes; triage memories is where the agent is told to come back and ask the human. The third step is materialising: a memory sitting in a store is useless if every call has to query it first, so a triaged learning gets written out to a skill, a coding standard, or the agent's own instructions file, where retrieval is free. His skills-versus-standards split is worth borrowing: a skill is something the agent knows how to do; a standard is a convention this project has agreed on and the model wasn't trained on. Retrieval is hybrid vector plus full-text with a three-stage walk — search, timeline anchor, then fetch full observations only for what's needed — and he was open that ranking hasn't been stressed yet; Nikolai pushed on exactly that, from indexing half a million Zulip messages, where finding everything is easy and ranking is the whole problem.