Home

tinyhivemind

TinyHiveMind provides coordination rules for agents that share a transcript. It resolves who belongs to a desk, who an authored mention addresses, what one participant may read, and who should respond. A host keeps the messages and decides when to call those rules.

The workspace also includes bounded group deliberation, semantic routing, and an OpenHuman adapter. Start with Architecture for the crate boundaries or Quick start for a runnable example.

Choose the part you need

If you need to...	Start with...
Resolve desks, rosters, mentions, and pure responder decisions	tinyhivemind-core and Desks and rosters
Read a host log, brief a seat, or maintain a bounded digest	tinyhivemind_core::runtime and Host integration
Divide work, track explicit completion, or deliberate by quorum	tinyhivemind_core::hive and Hive episodes
Route messages across host-neutral conversation surfaces	tinyhivemind_core::embed and Integration crates
Use TypeSafe System One for routing	tinyhivemind_core::typesafe and Integration crates
Bind agents to a desk and run completion episodes	tinyhivemind_core::driver, tinyhivemind-tools, and Integration crates
Run those episodes on OpenHuman agents	tinyhivemind-openhuman and Integration crates

How a turn works

A host appends a message with a global sequence. TinyHiveMind resolves its conversation and audience, chooses an eligible seat, and projects the part of the shared transcript that seat may see. The host runs the seat and appends its reply. Pure dispatch and referral rules can authorize a child turn when the host enables them.

A deliberation episode uses the same transcript. step can authorize a bounded round of seats. Blind seats may run concurrently because their views exclude their peers' current contributions. The host commits the returned episode state after every authorized turn in the round has been appended. Hive episodes describes the protocol; Episode policy covers its bounds.

For a group decision, agents can deposit typed markers such as !propose, !support, and !object in the shared transcript. The hive module reads those markers, gives recent and grounded contributions a score, and checks whether one proposal has enough distinct supporters. It can finish as converged, deadlocked, exhausted, or idle. The trace grammar and episode policy explain exactly how those outcomes arise.

The benchmark compares this protocol with a single responder and an independent vote given a matched budget. In one five-agent, four-option simulation, the tuned hive chose correctly in 82.1% of 5,000 seeded rooms, versus 78.5% for the vote and 57.6% for the single responder. Those are simulation results, not a claim that adding agents improves every real task. The benchmark report shows the controls and the limits of the measurement.

No library crate opens a database or owns the host's journal. The pure crates also have no runtime or transport dependency. Architecture explains those boundaries in detail.

Reading further

Transcript projection explains the shared log and viewer-specific histories. Shared medium covers what agents deposit in it. Benchmarks reports the measured results and their limits, with separate pages for scale, private exchange, and live runs. Glossary defines the terms used throughout the wiki.