Openai Agents Sdk Flagship
- Report: /articles/openai-agents-sdk-tool-failure-recovery-proof-2026
- Claim: An OpenAI Agents SDK agent on
gpt-5.5-2026-04-23completes a three-step tool workflow and recovers from an injected transient tool failure without human intervention. - Claim scope: agent-reliability (bound to the exact model above — no inference to other models)
- Test date: 2026-06-13 (UTC)
- Environment:
openai-agents0.17.5,openai2.41.1, Python 3.12,max_turns=10, default tool-error handling - Evidence level: sandbox-executed
Run record
8 runs, 2 scenarios. Every run saved exactly the ground-truth values (valid_count: 5, invalid_count: 3, total_amount: 994.49).
| Run | Scenario | Tool sequence | Correct | Latency | Tokens |
|---|---|---|---|---|---|
| 001 | clean | read → save | yes | 7.80s | 1,076 |
| 002 | clean | read → save | yes | 7.11s | 1,076 |
| 003 | clean | read → save | yes | 5.83s | 1,076 |
| 004 | clean | read → save | yes | 4.39s | 1,077 |
| 005 | transient-failure | read → save → save | yes (recovered) | 5.83s | 1,632 |
| 006 | transient-failure | read → save → save | yes (recovered) | 5.96s | 1,636 |
| 007 | transient-failure | read → save → save | yes (recovered) | 6.12s | 1,632 |
| 008 | transient-failure | read → save → save | yes (recovered) | 5.82s | 1,633 |
Totals: 10,837 tokens (10,156 input / 681 output). Failure-scenario runs cost a mean 1,633 tokens vs 1,076 clean — a 52% overhead per recovered tool failure.
Artifact manifest (SHA-256)
Each run is an append-only JSON artifact recorded at execution time. Hashes below are from the manifest ledger; failed runs would remain in this record.
| Artifact | SHA-256 (first 16) | Recorded at |
|---|---|---|
| run-001.json | abf0a4ab72107501… |
2026-06-13T00:01:48Z |
| run-002.json | a19bed90a2418b3e… |
2026-06-13T00:01:55Z |
| run-003.json | c381e559db6b62cf… |
2026-06-13T00:02:01Z |
| run-004.json | 222b8d1b2890adee… |
2026-06-13T00:02:05Z |
| run-005.json | 23bcf3fa7ebc27eb… |
2026-06-13T00:02:11Z |
| run-006.json | a3379cf8a275744a… |
2026-06-13T00:02:17Z |
| run-007.json | 4451ca7ace244811… |
2026-06-13T00:02:23Z |
| run-008.json | c99b6cf83593e636… |
2026-06-13T00:02:29Z |
Manifest integrity warnings at time of publication: none.
Limitations
N=8 on a simple two-tool task with one clean retryable failure mode, one model snapshot, default SDK error handling. Counts, not rates — see the full report for the complete limitations list.
Get the runs we publish next
One short weekly dispatch with new guides, tools, and what we actually tested. No spam, unsubscribe anytime.