The Protocol
A dual-process governance design for AI agent systems: generation stays focused on the task; critique intervenes only when needed.
Optimized for forward progress. It handles the primary task with as little overhead as possible, leaving the context window available for actual reasoning.
Mode: Generate. Ship. Iterate.
Optimized for critique, not generation. It looks for drift, repetition, and failure signals, and stays quiet unless intervention is warranted.
Mode: Observe. Detect. Intervene.
Current literature suggests language models are weak judges of their own reasoning without external feedback. Huang et al. (2023) and Zhang et al. (2024) both point in that direction. Janus takes that seriously by separating forward generation from critique instead of forcing both jobs into the same stream.
Governance should adapt to performance, not remain fixed at maximum intensity. The Janus Protocol starts with a simple heuristic and then adds richer signals when they are available.
First attempt. Trust is high. Agent operates freely. Governance is silent.
Second attempt, same problem. Environment stability check. Are we building on solid ground?
Third attempt, same pattern. Something is structural. Repeating is looping, not iterating.
Compare the current approach against recent attempts. High similarity can trigger escalation even when the retry is framed as something new.
Catches "same approach, different words"Infer confidence from the language of the session. Repeated hedging or sharp drops can trigger escalation without asking the agent to self-report.
Low reporting overheadThe original N-Pattern. A lightweight fallback when semantic or confidence signals are unavailable.
Minimum viable governanceA governance system is only as good as what happens next. These recovery paths are chosen to break loops, restore context, or bring a human in when the signal suggests it.
| Trigger | Recovery | Rationale |
|---|---|---|
| Semantic repetition | Alternative approach | Same approach is failing. Need fundamentally different strategy. |
| Confidence decay | Request human input | Agent uncertainty indicates information gap human can fill. |
| Iteration fallback | Decompose task | Problem may be too large. Break into smaller subproblems. |
| General | Reset context | Context window may be polluted. Fresh start with clean state. |
| General | Pause and reflect | Metacognitive intervention. Step back before stepping forward. |