Multi-Agent Coordination
Multi-agent deliberation grounded in shared memory. Conflict resolution. Bounded iteration. The substrate that makes multi-agent workflows actually agree on something.
The problem
Multi-agent setups silo. No shared context. No conflict resolution.
Most multi-agent setups silo. Each agent runs in its own LangGraph state, holds its own chain-of-thought, and emits its own conclusion. When two agents disagree, there's no protocol for resolution — usually the last agent to speak wins, or a "manager agent" picks a side without seeing the underlying evidence.
The result: multi-agent systems that look impressive in demos and produce inconsistent, ungrounded reasoning in production. Teams either drop back to single agents (losing the parallelism) or build their own coordination layer (badly).
How it works
Shared scratchpad in the memory fabric + deliberation primitives.
Cerebe's Multi-Agent Coordination puts the agents on a shared scratchpad in the memory fabric. Every agent reads + writes the same store, references the same evidence, and contributes to the same trace.
Deliberation primitives — propose, critique, defend, concede, escalate-to-human — let agents engage substantively rather than just running in parallel. Conflict resolution weights positions by confidence + evidence references. Bounded iteration (max rounds + consensus threshold + confidence floor) keeps deliberation from spinning. Each round is OTel-traced; the full transcript is stored in memory and replayable with updated context.
- Shared scratchpad lives in the memory fabric — every agent reads + writes the same store
- Deliberation primitives: propose, critique, defend, concede, escalate-to-human
- Conflict resolution: weighted by agent confidence + evidence references
- Bounded iteration: max_rounds + require_consensus + confidence thresholds
- Trace-correlated: each agent's contribution is a span in the same OTel trace
- Replayable: deliberation transcript stored in memory; can be re-run with updated context
# Multi-agent deliberation grounded in shared memory
from cerebe import Cerebe
cb = Cerebe(api_key=...)
debate = cb.agents.deliberate(
session="sess_123",
question="Is this patient ready to progress to phase-2 exercises?",
agents=[
{"id": "physio-agent", "persona": "movement specialist"},
{"id": "safety-agent", "persona": "injury-risk analyst"},
{"id": "compliance-agent", "persona": "patient-engagement tracker"},
],
max_rounds=3,
require_consensus=True,
)
# Each agent reads + writes to shared scratchpad (in memory fabric)
# {
# "rounds": [
# { "agent": "physio-agent",
# "position": "Yes — ROM improved 18° in 2 weeks, pain trending down",
# "evidence_refs": ["mem_a4b", "mem_c91"] },
# { "agent": "safety-agent",
# "position": "Hold — last 2 sessions reported minor flare; check recovery",
# "evidence_refs": ["mem_d22"] },
# { "agent": "compliance-agent",
# "position": "Yes — compliance trending up; progression is motivating",
# "evidence_refs": ["mem_e44"] },
# ],
# "consensus": "Hold for 1 week, then re-evaluate. Track recovery + flare.",
# "confidence": 0.78,
# "trace_id": "otel_a89...",
# }Get Started
Multi-agent reasoning that actually converges.
Shared scratchpad. Deliberation primitives. Conflict resolution. Bounded iteration. Trace-correlated. Full reference and self-signup at cerebe.ai.