Agents layer
v0.8BetaAgent Protocols
Coordination and tool-use patterns for reliable multi-agent systems.
- Python
- TypeScript
from legion_agents import Plan, swarm
plan = Plan.shared("multi-region-rollout")
async with swarm(agents=8, plan=plan) as fleet:
# Hand-off, arbitration, and recovery are protocol-level
await fleet.run(rollout_tasks)
- Version
- v0.8
- Status
- Beta
- License
- Open specification
- Implementations
- Python · TypeScript
- Stack layer
- Agents
What the spec defines
The guarantees, not the vibes.
Agent Protocols standardize how agents coordinate: task hand-off, shared planning state, and tool-use arbitration patterns that stay reliable as the number of agents grows. Most production agent failures happen at the seams — the protocols make each seam checkable.
The core insight from our coordination benchmarks: swap the models inside an agent stack and scores move modestly; swap unstructured seams for protocol-mediated ones and failure rates fall by multiples. Reliability is an interface property before it is a model property.
Foltrac's autonomous operations run on these protocols in production — multi-region rollouts coordinate in dependency order, and the rollback engine's recovery behavior is a protocol-level recovery hook, not bespoke incident code.
The Python and TypeScript SDKs implement the full protocol surface and interoperate out of the box. The spec is in beta at v0.8 while the arbitration section absorbs benchmark feedback; the coordination benchmark suite runs against any conformant stack.
Proven in production