Skip to content

tinyagents.md

Latest commit

 

History

History
50 lines (41 loc) · 2.48 KB

tinyagents.md

File metadata and controls

50 lines (41 loc) · 2.48 KB
Raw

TinyAgents

TinyAgents (inherited from OpenHuman at vendor/openhuman/vendor/tinyagents, crate tinyagents = "2.1", Rust 2024) is the recursive language-model harness: durable typed state graphs (checkpoints, interrupts, Send fan-out, subgraphs), a provider-neutral model harness (typed tools, middleware, structured output, streaming, usage/cost), a capability registry, the .rag/.ragsh surfaces, and an RLM runtime. It is embedded as a library behind the existing tiny feature — it is not a server.

What OpenCompany uses it for

  • StubBrain — the Phase 1 offline brain: a single harness call so the whole kernel pipeline is testable without the hosted backend (the mock provider makes this work in CI with no network).
  • Local worker execution — when the brain delegates sub-work that should run host-side (long tools, local file work), workers run as TinyAgents sub-agents under ToolProvider, with usage/cost rolled into the ledger.
  • Built-in tools — the fallback ToolProvider when OpenHuman is absent.
  • SidecarBrain inference proxy — the sidecar's InferenceClient calls back into the host, which fulfils it through the TinyAgents harness.
  • Observability — the embedded Langfuse exporter, proxied through the TinyHumans backend with the same credential (runtime/config.md).
  • NativeBrain (far future) — a graph port of Medulla's runCycle; interface reserved, no commitment (medulla.md).

Roster → execution mapping (internal detail)

A manifest [[agent]] entry does not become a standing process. The Roster is charter data the brain reasons over; when a cycle delegates, host-side workers are ephemeral TinyAgents sub-agents parameterized by the teammate's role/description/grants, with recursion depth and budget caps enforced by the harness (SubAgentDepth, usage rollup). This mapping is never surfaced to prosumers (glossary, translation table).

Registry use

The TinyAgents registry gives names to capabilities (models/tools/agents/ graphs) inside a company runtime; the kernel registers the granted tool catalog and worker archetypes there so .ragsh debugging and future .rag blueprints see the same catalog the brain does.

Boundary

TinyAgents never talks to the network on its own in OpenCompany: model access goes through the TinyHumans credential; provider keys beyond that are out of scope (roadmap.md, non-goals — not a model host).