← @bgSponge

Agent Swarms are a Distributed Systems Problem

Summary

Chroma's Foundation memory layer has an agent swarm ingest traces and company data into a shared wiki.

  1. Agent work is too expensive to abort.

    A database retry re-runs the same code in milliseconds. An agent re-discovers its read set through search and reasoning, which costs minutes and tokens.

    when a database transaction aborts, retrying re-runs the same code in milliseconds.

  2. Atomicity across the wiki is a non-goal.

    A valid reasoned edit is cheaper to leave than revert. Individual page writes stay atomic through Chroma Cloud optimistic concurrency control.

    the swarm is editing one shared state, just with the collisions postponed.

    atomicity across the wiki as a whole is an explicit non-goal of Foundation.

  3. Fission is two-phase locking plus early commit.

    Exclusive locks at first page read, not over search results; wound-wait on batch arrival timestamps; no rollback. Abort leaves committed work.

    Aborting is treated as early commit, not roll back.

  4. Git merge throws away finished work.

    An early Foundation used git; three of eight calls gave up after conflicts and four after stale reads. Agents skip the contended page instead of resolving it.

  5. Prefix-safe writes make anomalies benign.

    A TODO tool keeps every write prefix valid. Haiku 4.5 plus Context-1 was enough to retry from observed state; 39.7% of wounded read-waits revisited the wounding page.

Read the original ↗Save to your Sponge

Summarized with Sponge. Quotes are the author’s own words.

More from @bg

All readings →