Skip to content

Latest commit

 

History

History
49 lines (43 loc) · 2.83 KB

File metadata and controls

49 lines (43 loc) · 2.83 KB

Research

Where the mechanisms in this workspace come from, at more depth than ../../wiki/Further-reading.md carries.

The wiki page is a reader's map: one paragraph per mechanism, linked to the page that uses it. These files are the working notes behind it — the equations, the measured constants, the exact citation, and, for each mechanism, the one line that matters here: what a shared transcript would have to represent to implement it.

A note is worth writing when a design decision needs a warrant that is longer than a comment and is not itself a decision. Decisions go in ../adr/; behavior goes in ../specs/; what actually happened when it was run goes in ../experiments/. These files hold only the reading.

Notes

  • biology.md — stigmergy, quorum sensing, honeybee nest-site selection and its differential equations, response thresholds, and the limits of collective intelligence.
  • shared-context.md — the human and organizational half (transactive memory, distributed cognition, grounding, boundary objects, awareness) and the open-source landscape of shared agent memory.
  • delegation.md — how a collective decides which specialist acts: response thresholds, the tremble dance, transactive memory, hidden profiles, and the 2023-2026 landscape of agentic routers, with the honest half of the evidence against each.
  • long-context.md — position bias in a long window, context rot, and recursive language models: why P14 makes the transcript queryable instead of making the window bigger.
  • context-in-agent-teams.md — what it costs when two agents on one desk hold different transcripts: the Cognition/Anthropic disagreement about sharing context, hidden profiles, the conformity that full visibility buys, the auditability a private channel owes, and the four ways a divergent view fails a reader with a sliding window. The reading behind P17.
  • multi-context.md — N seats, N windows, one record: why no shipping design gives its orchestrator the full context, what a seat carries between turns instead (context folding, externalized plans), and the feedthrough a shared artifact already emits. The reading behind seat continuity and ../specs/thoughts-and-channels.md.
  • grok-bots/ — twelve notes on the open-source Grok Bot ecosystem, read at pinned commits: how each one models a roster, what makes a message start a turn, and how four of them shipped uncontrolled fan-out and then bought a bound.

Each closes with a table mapping the mechanisms it covers to the state this workspace already holds, and to the state it does not.