Agent Workspace Practice

Where it sits in the directory's order

What can be seen from outside, weighed as the directory publishes.

Where it answers

• https://jamesross.ai/mcpstreamable-http

The tools it lists (6, as of 21 hours ago)

The tool-description rules found nothing in these descriptions. They look for instructions aimed at a model and for hidden characters; they cannot see what a tool does when it runs.

• search_architecture

  Full-text search across the published architecture corpus: the 18 patterns and how they compose, the evaluation method, the seven learn-track modules, and the teardowns of other people's published agent systems. Returns ranked sections, each with its title, its source file, a snippet, and a link to the full text on GitHub. Use it when you want the passage itself rather than a pattern summary; lookup_pattern is the narrower tool that answers "which pattern covers this?". Every section is also readable in full as a resource under the architecture:// scheme (resources/list, resources/read).

• lookup_pattern

  Search the 18 architecture patterns behind this practice by keyword or by pattern number. Returns up to 3 matches, each with its number, title, the problem it opens on, and a link to the full text. Use it when you want the reasoning behind an agent-workspace design decision rather than a how-to.

• assess_workspace

  Score an agent workspace across six dimensions: canonical knowledge, memory and context economics, verification and oversight, safety and permissions, telemetry and cost, provenance and delegation. Two variants: "practitioner" takes 18 answers (a single operator's workspace), "org" takes 12 (an organisation's agent readiness). Each answer is an integer 0-3 in question order, where 0 is the least mature option offered and 3 the most. Returns per-dimension scores, an overall percentage, a maturity band, and the two weakest dimensions with the gap that matters there. The question text lives at ht

• price_context_read

  Price a block of context by where it enters a session. The API is stateless, so the whole transcript is re-sent on every inference step, and a token costs its size times the number of turns left after it arrives. Returns the carried token volume, the cost of the block in USD, what the same read would have cost entering 20 turns from the end, and the multiplier between the two. Use it to decide whether a large read should move later in a session or into a subagent.