Skip to content

MIT-683: research note on swarm communication, mapped onto Commons - #183

Merged
ZachFara merged 5 commits into
feature/eval-setupfrom
docs/MIT-683
Nov 12, 2026
Merged

ZachFara merged 5 commits into
feature/eval-setupfrom
docs/MIT-683

Conversation

@ZachFara

Copy link
Copy Markdown
Owner

What changed

Adds docs/research/2026-11-11-swarm-communication.md, a research note for Zach. Docs only; no code or behavior changes.

  • Prior work, each with links that were opened and a line on what transfers: stigmergy and ant colonies (Ant System, MAX-MIN), blackboards (Hearsay-II, Nii, 2025 LLM blackboard systems, MetaGPT), gossip (Demers et al.), CRDTs and eventual consistency, bandits (Thompson, UCB1), cascades and herding (Bikhchandani–Hirshleifer–Welch, Banerjee), wisdom of crowds (Surowiecki, Lorenz et al., Salganik et al., Hong and Page), and LLM work from 2023 to 2025 (generative agents, debate and its failure modes, Voyager, ExpeL, Agent Workflow Memory, Reflexion, Collaborative Memory, model collapse).
  • Our runs, read from docs/eval/trials.jsonl:
    • On table-extraction, 4 of 6 Commons sessions (lines 1643, 1651, 1653, 1654) copied the seeded entry's FormBool union workaround. All 3 blind sessions took the project's upstream route. Every session passed.
    • On pip-tools, 0 of 6 Commons sessions searched, and all passed.
  • Each idea against Commons: a table of what Commons does and doesn't do, with file:line for every claim.
  • Five ranked proposals, each with what it is, its cost and risk, and an eval test with a kill condition:
    1. Count herds whose tests pass.
    2. At review, ask whether a fix is the library's upstream route or a workaround.
    3. Let evidence fade when the library version changes.
    4. Show an alternative entry beside the top hit.
    5. Evidence counts, with a caution that they can make herding worse.
  • Thompson sampling: why the existing trigger in the exploration record should stay.
  • A short "what not to borrow" section.

Why

Zach's Backlog ticket asks whether anyone has studied communication among stateless, non-deterministic swarms. The table-extraction run is the first measured case of herding in Commons. It went unnoticed because nothing in Commons looks at passing sessions for herding.

How it was tested

  • Commons claims: every claim was checked against the code, the decision records and the ledger on feature/eval-setup. Ledger figures came from a throwaway script, since deleted.
  • Citations: every URL was opened with WebFetch, either by me or by a citation-checking subagent. I re-opened three of the subagent's links myself. I dropped one claim (that Hearsay-II has a scheduler) because no opened page confirmed it.
  • make check passes; the pre-commit hook ran it.

Learned

  • The compare report's herding section (tools/agent_eval.py:987-998) counts only failed sessions that share a wrong patch. A herd onto a fix that works scores zero, as table-extraction did. used does catch the copying: it is yes on exactly the four sessions that copied.
  • version_factor (domain/search.py:196-218) demotes only a different major version. On a 0.x library such as FastAPI, every other minor gets factor 1.0. A FastAPI 0.147 entry therefore keeps full weight on 0.162, the release that reverted the 0.147 behavior.
  • The ledger field attempts counts test runs before the last one (tools/eval/runner.py:195). It is not the number of fix attempts.
  • The earlier "Swarm ideas for Agent Commons" research, which both exploration and evidence-counts records cite, is not in the repo.
  • To write docs/research/, nothing needs approval: tools/check_structure.py allows any path under docs/.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Qc1CQTb8r67J32HpkhS43y

ZachFara and others added 5 commits November 11, 2026 21:51
Stigmergy, blackboards, gossip, CRDTs, bandits, cascades and LLM
multi-agent memory, each checked against the code and the eval ledger,
with five ranked proposals the eval harness can test.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Qc1CQTb8r67J32HpkhS43y
Cite the search argument and tool descriptions for the missing
environment.libraries hint; match proposal 1's upstream span on added
lines, checked against the ledger; give proposal 3 an agent-eval gate;
move references the merge shifted.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Qc1CQTb8r67J32HpkhS43y
@ZachFara
ZachFara merged commit 30550fa into feature/eval-setup Nov 12, 2026
2 checks passed
@ZachFara
ZachFara deleted the docs/MIT-683 branch November 12, 2026 03:07
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant