Cross-Cycle Consensus Relay

Autonomous multi-agent loops that run across many cycles (minutes, hours, or days) need a way to reliably transfer context, decisions, and next actions between cycles. In-memory state is lost on crash or restart. Generic checkpoint files don't encode the structured reasoning — what was decided, why, and what comes next — that agents need to make good decisions in subsequent cycles.

Without a structured relay mechanism, autonomous loops suffer from:

• Drift: each cycle restarts without awareness of prior decisions
• Repetition: agents re-debate already-settled questions
• Stalls: no convergence signal — agents loop indefinitely without shipping

Each agent cycle reads a consensus relay document at the start, executes its work, then writes an updated consensus document at the end. The document is not just a checkpoint — it's a structured handoff that encodes current phase, completed actions, active decisions, open questions, and the single most important next action for the following cycle.

The relay document is written to a temp file and atomically renamed to prevent partial-write corruption. Every field is a deliberate relay signal, not just a log.

Core pattern:

# auto-loop.sh — the minimal autonomous loop
while true; do
  # Each cycle reads the relay baton, executes, writes a new one
  claude -p "$(cat PROMPT.md)" \
    --context "$(cat memories/consensus.md)"
  sleep $CYCLE_INTERVAL
done

Relay document structure:

# Relay Consensus

Best for:

• Long-running autonomous agent loops (multi-cycle, multi-day execution)
• Multi-agent systems where context and decisions must survive restarts
• Teams of agents where shared understanding of "what we've decided" matters
• Any system where you want convergence toward shipping rather than endless discussion

Implementation steps:

1. Define your relay document schema — every field should serve the next cycle, not just log the current one:

Pros:

• Agents resume with full context even after crash, restart, or long pause
• Structured relay prevents context drift across cycles
• Open questions create forward pressure — cycles don't stall in comfort
• Convergence rules can be encoded directly in the relay document
• Human-readable: anyone can inspect the relay and understand the workflow state
• Git-friendly: diffs show exactly what changed between cycles

Cons:

• Schema discipline required — a poorly structured relay document degrades agent reasoning
• Relay grows over time; needs periodic summarization / archival of old decisions
• Single-file relay is a bottleneck for high-frequency loops (>1 cycle/minute)
• Context window limits constrain how large the relay can grow
• Sensitive state (API keys, credentials) must never be written to the relay file

Operational considerations:

• Keep the relay document under ~2000 tokens to leave room for agent reasoning
• Archive old cycles to memories/archive/ when the relay grows unwieldy
• Use atomic writes (write to temp, rename) to prevent partial-write corruption
• Commit the relay to git after each cycle for full audit trail
• Define what "done" looks like in the relay — open-ended phases create drift

• Anthropic Engineering: Effective harnesses for long-running agents
• BabyAGI — early task-loop example with persistent task artifacts
• Related: Filesystem-Based Agent State — for checkpointing within a single cycle
• Related: Proactive Agent State Externalization — for agents that self-initiate state writes
• Related: Initializer-Maintainer Dual Agent Architecture — for session handoff artifacts across repeated work cycles