Add: cross-cycle-consensus-relay

[NikitaDmitrieff]

New Pattern: Cross-Cycle Consensus Relay

Category: Orchestration & Control

What it is

A structured markdown document that each autonomous agent cycle reads at the start and writes at the end — acting as the state relay between cycles. Not just a checkpoint file; it encodes current phase, completed actions, key decisions (with reasoning), active projects, and — critically — the single next action for the following cycle.

Why it's different from Filesystem-Based Agent State

The existing Filesystem-Based Agent State pattern addresses resumption after failure and long-running task checkpointing. This pattern addresses the multi-cycle relay problem: how do you transfer structured reasoning, institutional decisions, and forward directives between autonomous cycles that may run hours or days apart?

Key differences:

• Relay, not checkpoint — designed to be read by the next cycle, not just to resume the current one
• Convergence rules — includes stall detection ("same next action 2 cycles = change direction")
• Phase-aware — encodes where the project is in its lifecycle, not just transactional state
• Human-readable artifact — the relay document is the "institutional memory" of the autonomous company

Real-world backing

This pattern is used in production by auto-co, an open-source autonomous AI company framework that has run 13+ cycles building its own infrastructure and landing page. The memories/consensus.md file is the relay baton for the entire company.

The atomic write protocol (write to .consensus.tmp, then mv) prevents partial-write corruption in long-running loops.

Fits your "repeatable" criterion

Multiple teams building autonomous agent loops independently arrive at this pattern. The core insight — that you need a structured relay document, not just generic state files — is not obvious and worth cataloguing.

---

Continue Tasks: ❌ 3 failed — View all

[vercel]

@NikitaDmitrieff is attempting to deploy a commit to the nkkko's projects Team on Vercel.

A member of the Team first needs to authorize it.