Shipped: One Deny, Then the Current Version
In agent-coherence 0.16.1 and plugin 0.7.1, a strict-mode Claude Code session whose copy of a file went stale is denied once, and its next Read gets the current version.
Read the post →Blog
Engineering notes on shared state in multi-agent systems — stale reads, lost updates, and what it takes to make agent memory safe to write to.
In agent-coherence 0.16.1 and plugin 0.7.1, a strict-mode Claude Code session whose copy of a file went stale is denied once, and its next Read gets the current version.
Read the post →In v0.16.0 a writer that opts in is refused a file another agent holds, instead of taking it, and a holder gives the file back when it is done.
After Claude Code compacts a session, the summary can drop which files it held and which ones a peer changed. A hook can hand that record back.
MCP removed sessions by design and its roadmap names no shared-state work. If the protocol is to be stateless, coherence has to live in the layer above it.
A 1,200-agent swarm reinvented locks, votes and successors. Why those primitives break for agents, and what cooperative, non-blocking work and handoff need.
In v0.15.0 an agent can hand a file to a named successor, and on a single host the Python coordinator refuses the giver's own writes until the handoff ends.
One record your infrastructure can verify, one the model narrates. They agree until something goes wrong, and the narrated one is the one you check first.
v0.14.0 brings workspace checkpoints with per-member honesty, an effect fence on the MCP surface, and a session that re-anchors after compaction.
Your verifier approved a clean answer that was still wrong, because it was the right answer to a stale question. Output checks never look upstream.
A lock, a lease, and a fencing token answer three different questions. The write your version check passes is the one the fence refuses.
When an update vanishes the reflex is a store migration. The store was never the problem. The fix is the version on the write, over the store you already run.
The failures arrive in a fixed order as an agent workspace grows: stale reads, then lost updates, then zombie writes. Find your rung and see the next one coming.
BYO-substrate bindings are live in v0.13.0. CoherentRow guards a Postgres row, CoherentObject guards an S3 object, and the coordinator holds a version, MESI state, and a content hash, never your bytes. A peer's commit marks your cached read stale before you act on it, on a single host, with the same typed conflict and the same reacquire() recovery on every substrate.
The coordination lever is live. Version-CAS, a read-generation fence, and MESI invalidation-deny stop the concurrent lost-update and the sequential stale-read-then-write on a single host, with five TLA+ invariants checking they hold.
The vanished write is not one bug. It is a class with five members, and prevention is not one lever. A durability receipt, a retention policy, and a coordination primitive split the five between them, and only the last one is the runtime's to ship.
Thirty agent-memory products describe how agents remember. Not one describes what happens when two writers hit the same key. The lost update is the half that breaks production.