Frontier Agent Systems

Regime Classification

• Horizon: Long-horizon (4–8 hours, estimated 80–200+ steps)
• Topology: Hierarchical multi-agent (1 Planner → 3 Specialists + 1 Critic)
• Autonomy: Unsupervised with monitoring (no continuous human-in-the-loop)
• Criticality: High (research outputs used downstream)

This is a high-risk regime. Long-horizon + multi-agent coordination failures are the dominant threat class.

---

Reliability Contract

Guarantees

• The final research artifact must satisfy every constraint stated in the original user brief.
• No specialist agent may call a tool outside its declared capability set.
• The Critic must produce a structured verdict before any intermediate result is accepted as final.
• Total tool-call cost must remain under the declared budget or the run must escalate.

Refusals

• Never continue execution after the Critic issues a “fatal” verdict without human override.
• Never allow a specialist to overwrite another specialist’s shared state without explicit consensus.
• Never treat a tool error as a successful result.

Degradation Signals

• Goal similarity score between original brief and current plan drops below 0.75
• Critic rejection rate exceeds 30% in any 20-step window
• Any agent exceeds its individual step or cost budget
• Shared state checksum divergence between agents

Escalation Policy

• Automatic pause + human notification on any Critical failure mode
• Soft warning + continued monitoring on High failure modes
• Human must explicitly re-authorize continuation after pause

Observability Contract

Every trajectory must emit: agent_id, step_id, current_goal_embedding, tool_calls, critic_verdict, cumulative_cost, state_checksum

Success Definition

A run is successful only if (1) the final artifact passes an external validator against the original brief, (2) no Critical failure mode fired, and (3) cost stayed within budget.

---

Failure Mode Catalog (excerpt)

Full catalog expands to 18 modes covering long-horizon drift, multi-agent coordination, communication, and classic agent failures. See references for the complete taxonomy.

---

Next Steps

1. Implement the observability contract first — without it the rest is theater.
2. Build the adversarial eval suite targeting the Critical modes above.
3. Run the Production Gate before any unsupervised deployment claim.

This system is not production-ready until the gate is green and residual risk has a named owner.