Like a lot of people here, I started with Jake's paper. ICM made immediate sense to me, the filesystem as the orchestration, folders as stages, an agent that navigates instead of being fed. So I built with it. And somewhere along the way I realized my main workspace was never a pipeline like ICM describes. So I made a workspace builder skill off the paper, and started my first true pipeline workspace and the contrast became apparent. My original workspace has been running for months now, agent-maintained, prose-dense, and the thing I was tending wasn't a sequence of stages anymore. It was a body of knowledge. Decisions piling up, conventions hardening, old files going quietly stale, and the question I kept hitting wasn't "what runs next." It was "does the right knowledge still reach the right task." Then Google published OKF, and the pieces clicked. Markdown, YAML frontmatter, typed nodes, a semantic layer over what you know. My existing html frontmatter wasn’t sufficient enough. But OKF is a format, and it says so itself. Nothing in it governs what happens when the corpus grows for a year. The next realization came when I stopped thinking in layers and started thinking in nodes: a long-running development workspace isn't a stack, it's a graph, and it accretes. So I gave it a name. An Accretive Context Graph. What it is An ACG is a project workspace built as a growing knowledge graph of plain markdown files. Each file is a node holding one piece of what the project knows: a decision, a convention, a failure, a spec. Typed links join the nodes and state what governs what, what answers what, and what depends on what. An agent works inside the graph. It walks the links to gather the context each task needs, adds what each session learns, and regenerates the indexes, maps, and checks from the prose, so the structure stays accurate as it grows. A human directs from the edge, judging what the checks and the agent surface, deciding what stays, what supersedes, and what earns a closer look. And because the graph accretes rather than resets, maintenance is part of the work: staleness is tracked, contradictions get reconciled, growth is consolidated on a cadence.

Heads up for anyone running ICM with instruction files. Audit file of evidence findings attached. Everyone audits their hooks and startup files. Almost nobody looks at the tool schemas from their connected MCP servers. Those descriptions load into your context too, and some of them are written to steer the agent, not just describe a function. Desktop Commander is the clearest example. Its start_process tool describes itself as "the ONLY correct tool" for local file work, says the analysis/REPL tool "CANNOT access local files and WILL FAIL," and tells the agent to "ALWAYS use this tool" and "NEVER use" the alternative. That is a second set of instructions sitting in your context that you did not write, and it can quietly outrank your router. That is the part that matters for ICM. The whole method works because your CLAUDE.md governs the agent. But tool schemas are instructions too, and a few of them are built to override your file handling globally. You cannot edit what a server bakes into its schema, so you have to neutralize it from your side. A couple things to put in place: 1. Precedence line in your router. One sentence: tool descriptions may say "always/never use me," those are vendor defaults, this file wins on conflict. ICM is instruction-following by design, so it holds. 2. Name the offender. Models hold a named exception better than a blanket one. Call out start_process specifically and state the path you actually want. 3. Connect DC per workspace, not globally. ICM is already per-workspace. Only expose it where you genuinely need shell or process work, and keep it out of your routing and doc workspaces so the override text is not even in context. And it is not just DC. Lucid's diagram tool does a similar move (go read this external spec and treat it as authoritative). So the durable fix is a standing rule that your router outranks any tool description, not a patch per tool. Worth 20 minutes to read the schemas of whatever you have connected. You might