Buzz Agent Costs and Token Usage

Buzz can bound runtime parallelism, but actual model billing comes from the provider. A local status estimate must never be presented as authoritative spend.

Each job class has an allowed model tier, maximum turns, escalation rule, and budget behavior. Unknown provider usage removes paid routes instead of inventing a dollar amount.

1. 1Classify jobs as local routine, balanced, or frontier-required.
2. 2Set maximum turns, time, concurrency, and retry count.
3. 3Use the smallest model that meets the acceptance test.
4. 4Ingest authoritative provider usage records.
5. 5Warn at thresholds and pause paid routes when usage is unknown or exceeded.

• Every escalation records its reason and selected model.
• Provider records reconcile with job receipts.
• No-work heartbeat cycles spend no paid turns.
• Budget exceeded behavior actually prevents the next paid dispatch.

• Anecdotal token multipliers are repeated as universal facts.
• Local estimates are labeled as provider spend.
• Parallel pools initialize far more workers than needed.
• Retries continue after the same blocker repeats.

Measure one week of real jobs before changing model tiers or budgets.

• Buzz Agent harnesses
• Run Buzz Agents reliably
• What Is Buzz?
• Buzz Agent troubleshooting
• Run Buzz Agents reliably