Day 53: Avoiding Agent Ping-Pong

Two agents politely disagreeing for forty rounds is not “reasoning” — it is a token furnace. Production swarms need budgets, stop conditions, and a human gate after N rounds so ping-pong cannot become the default control flow.

⚡ TL;DR: Cap rounds, tokens, and wall-clock per run. Detect oscillation (repeated REJECT reasons). After N critic failures, route to a human with a one-page brief — never auto-loop overnight.

Budgets as first-class state

# runtime/budget.py
from dataclasses import dataclass

@dataclass
class RunBudget:
    max_rounds: int = 4
    max_tokens: int = 200_000
    max_wall_s: int = 900
    tokens_used: int = 0
    rounds: int = 0

    def charge(self, tokens: int):
        self.tokens_used += tokens
        if self.tokens_used > self.max_tokens:
            raise RuntimeError("budget_tokens_exhausted")

    def tick_round(self):
        self.rounds += 1
        if self.rounds > self.max_rounds:
            raise RuntimeError("budget_rounds_exhausted")

Detect oscillation

def oscillating(history: list[str]) -> bool:
    if len(history) < 4:
        return False
    # ✅ same reject fingerprint twice in a row pair
    return history[-1] == history[-3] and history[-2] == history[-4]

When oscillation trips, stop — do not “try a stronger model” in a loop without a new plan hash.

Human after N rounds

// gates/human.ts
export type Gate = { kind: "continue" } | { kind: "human"; brief: string };

export function afterCritic(rounds: number, verdict: { decision: string; evidence: string[] }): Gate {
  if (verdict.decision === "ACCEPT") return { kind: "continue" };
  if (rounds >= 3) {
    return {
      kind: "human",
      brief: `Stuck after ${rounds} rounds.\nEvidence:\n- ${verdict.evidence.join("\n- ")}`,
    };
  }
  return { kind: "continue" };
}

❌ “Keep going until ACCEPT” with no human path — that is how weekend bills appear.

Closing checklist

• [ ] Enforce round + token + wall-clock budgets
• [ ] Hash reject reasons; detect oscillation
• [ ] Escalate to humans with a structured brief
• [ ] Persist budget counters in the run record
• [ ] Alert when runs hit caps frequently

Series navigation

Day 52: Blackboard vs Message Bus Orchestration · Day 54: Specialist Tools per Agent

Last updated September 11, 2026