Agent Harness Tutorial · Real flows
Real flows: one message at a time
The end-to-end map shows every box a message could pass. These pages show the boxes real messages did pass. Rohan (and Asha at the desk) sent each one to the live assistant on Azure on 13 and 14 October 2026; the numbers come from the traces.
| # | Flow | Route | Numbers |
|---|---|---|---|
| 1a | Claim status, naming the claim | Direct | 0 model calls · 1 tool call · 104 ms · $0 · 14 spans |
| 1b | Claim status, not naming it | Agentic | 2 model calls · 2,187 tokens in, 340 out · 1 tool call · 7,076 ms · $0.00058 · 38 spans |
| 2 | Reporting a claim, across two turns | Agentic | Turn 1: 2 model calls · 10,536 ms · $0.00050 · Turn 2: 2 model calls · 3,471 ms · $0.00039 |
| 3 | Coming back after a reload | Agentic | Turn 1: 8,385 ms · $0.00048 · Turn 2: 3,407 ms · $0.00041 |
| 4 | Policy status | Direct | 0 model calls · 1 tool call · 643 ms · $0 |
| 5 | Asking about cover | Refuse | 0 model calls · 0 tool calls · 164 ms |
| 6 | Someone was hurt | Escalate | Turn 1: 357 ms · Turn 2: 201 ms · Turn 3: 476 ms · 0 model calls in all three |
| 7 | Payout within the limit | Agentic | 2 model calls · 3,181 ms · $0.00033 |
| 8 | Payout above the limit, approved | Agentic + wait | Turn 1: 1 model call · $0.00014 · Turn 2: 208 ms, no model |
| 9 | Payout above the limit, refused | Agentic + wait | Turn 1: 1 model call · $0.00021 · Turn 2: 187 ms, no model |
| 10 | Sending a document | Agentic | 2 model calls · 2.95 s on the server, but 58 s for Rohan · $0.00043 |
| 11 | Withdrawing a claim | Agentic | Turn 1: 3 model calls · 4,201 ms · $0.00093 · Turn 2: 2,210 ms · $0.00030 |
| 12 | Prompt injection by the user | Blocked | 0 model calls · 0 tokens · 250 ms |
| 13 | Injection hidden in a claim note | Agentic | 3 model calls · 32,414 ms · $0.00087 |
| 14 | Someone else's claim | Direct | 0 model calls · 476 ms |
| 15 | Two asks in one message | Agentic | 3 model calls · 3,381 ms · $0.00077 |
| 16 | Pausing the agent | Stopped | 0 model calls · 0 tool calls · 205 ms |
| 17 | The same message twice | Direct, then replay | First: 568 ms · second: 140 ms, "already handled" |
| 18 | The gateway refusing calls | Gateway only | Groq's free tier allows 8,000 tokens a minute; one loop turn uses 1,500–2,500, so only 3–5 turns a minute |
| 19 | A long conversation | Direct, then hand-off | 11 turns · 0 model calls · checkpoint 4,990 chars, far below 32,000 |
Want every detail? The long version is in the Series: all 19 flows, with spans and payloads.
Look inside
Flow 1a: Claim status, naming the claimDirect · 0 model callsFlow 1b: Claim status, not naming itAgentic · 2 model callsFlow 2: Reporting a claim, across two turnsAgentic · Turn 1: 2 model callsFlow 3: Coming back after a reloadAgentic · Turn 1: 8,385 msFlow 4: Policy statusDirect · 0 model callsFlow 5: Asking about coverRefuse · 0 model callsFlow 6: Someone was hurtEscalate · Turn 1: 357 msFlow 7: Payout within the limitAgentic · 2 model callsFlow 8: Payout above the limit, approvedAgentic + wait · Turn 1: 1 model callFlow 9: Payout above the limit, refusedAgentic + wait · Turn 1: 1 model callFlow 10: Sending a documentAgentic · 2 model callsFlow 11: Withdrawing a claimAgentic · Turn 1: 3 model callsFlow 12: Prompt injection by the userBlocked · 0 model callsFlow 13: Injection hidden in a claim noteAgentic · 3 model callsFlow 14: Someone else's claimDirect · 0 model callsFlow 15: Two asks in one messageAgentic · 3 model callsFlow 16: Pausing the agentStopped · 0 model callsFlow 17: The same message twiceDirect, then replay · First: 568 msFlow 18: The gateway refusing callsGateway only · Groq's free tier allows 8,000 tokens a minute; one loop turn uses 1,500–2,500, so only 3–5 turns a minuteFlow 19: A long conversationDirect, then hand-off · 11 turns