Research investigation R0911 / comparison

The Managed-Agent Compliance Gap: Where OpenAI’s New Agents API Can Actually Run

At launch, the managed harness supports customer-controlled execution through self-hosting, but the OpenAI-managed control plane keeps agent sessions non-ZDR and U.S.-resident. That leaves no sampled configuration documented to combine ZDR, full-stack non-U.S. residency, retained sessions, and managed file-capable execution.

Current public editionSep 12, 2026Sep 12, 2026
Verified observations
18

16 measured fields

Supported claims
9

9 material findings

Cited sources
12

12 primary or authoritative

Research score
86

Automated topic and evidence score

Interactive figureThe Managed-Agent Compliance Gap: Where OpenAI’s...
CSV JSON
Data status9 verified records across 1 period

Snapshot only. There is not enough history to claim a trend yet.

Verified observationHover or focus any mark for exact valuesLast updated Sep 12, 2026
Coverage note

This synthesis uses the supplied public evidence available by September 11, 2026. The sample covers 11 documented execution configurations and excludes private contractual controls and provider enterprise features not publicly documented.

Dataset ID
spd:the-managed-agent-compliance-gap-where-openai-s-new-agents-api-can-actually-run-55aac41e
Stable URL
/research/the-managed-agent-compliance-gap-where-openai-s-new-agents-api-can-actually-run-55aac41e
Version
Live
Coverage
2026-09-11
Records
18
Fields
7
Updated

Read the data

The records behind the figure

CSV JSON
The Managed-Agent Compliance Gap: Where OpenAI’s New Agents API Can Actually Run data records
EntityMetricValueUnitObservedSourceTransform
/v1/agentsapplication-state retentionuntil deleted2026-09-11https://developers.openai.com/api/docs/guides/your-data
/v1/filesapplication-state retentionuntil deleted or configured expiry2026-09-11https://developers.openai.com/api/docs/guides/your-data
Launch compatibility sampleconfigurations requiring a third-party sandbox provider9 of 11configurations2026-09-11https://developers.openai.com/api/docs/guides/agents-api/environments/self-hostedExcluded OpenAI-hosted and generic customer-managed self-hosted execution.
Generic self-hosted executioncustomer-VPC executionsupported2026-09-11https://developers.openai.com/api/docs/guides/agents-api/environments/self-hosted
OpenAI-hosted sandboxcustomer-VPC executionincompatible; use self-hosted for a private network2026-09-11https://developers.openai.com/api/docs/guides/agents-api/environments/openai-hosted
/v1/agentsdefault abuse-monitoring retention30 daysdays2026-09-11https://developers.openai.com/api/docs/guides/your-data
Agents API sessiondeletion completionAPI removal immediate; physical cleanup may continue asynchronously2026-09-11https://developers.openai.com/api/docs/guides/agents-api/sessions/manage
Launch compatibility sampleexecution configuration count11configurations2026-09-11https://openai.com/index/introducing-the-agents-apiCounted OpenAI-hosted, generic self-hosted, and nine named provider integrations.
Launch compatibility samplefull-stack non-U.S.-residency configurations0 of 11configurations2026-09-11https://developers.openai.com/api/docs/guides/agents-api/overviewCounted only configurations regionalizing both OpenAI control-plane state and execution.
OpenAI-hosted sandboxidle deletion threshold1 hour after activity and keep-alives stophours2026-09-11https://developers.openai.com/api/docs/guides/agents-api/environments/openai-hosted
Launch compatibility sampleOpenAI Artifacts API support1 of 11 configurationsconfigurations2026-09-11https://developers.openai.com/api/docs/guides/agents-api/environments/filesCounted OpenAI-hosted as supported; generic self-hosted and nine provider configurations use provider or mounted storage.
OpenAI-hosted artifactpost-sandbox lifetimesurvives sandbox expiry; exact maximum unspecified; deletable2026-09-11https://developers.openai.com/api/docs/guides/agents-api/environments/files
Named sandbox providersprovider count9providers2026-09-11https://developers.openai.com/api/docs/guides/agents-api/environments/self-hostedCounted providers in the documentation table.
Nine named providersprovider-specific VPC mapping9 unspecifiedproviders2026-09-11https://openai.com/index/introducing-the-agents-apiCompared the aggregate VPC statement with the provider setup guides; none mapped VPC support by provider.
Self-hosted environmentsession deletion effect on computedoes not stop provider compute; separate cleanup required2026-09-11https://developers.openai.com/api/docs/guides/agents-api/environments/lifecycle
OpenAI-hosted sandboxworkspace-file lifetimeacross turns while the sandbox exists2026-09-11https://developers.openai.com/api/docs/guides/agents-api/environments/openai-hosted
Launch compatibility sampleZDR-compatible configurations0 of 11configurations2026-09-11https://developers.openai.com/api/docs/guides/agents-api/overviewApplied the documented Agents API-wide ZDR restriction to all sampled environment choices.
Launch compatibility sampleZDR-incompatible configurations11 of 11configurations2026-09-11https://developers.openai.com/api/docs/guides/your-dataApplied /v1/agents ZDR ineligibility to all 11 configurations.

Measurement technique

How to read this report

  1. 01Evidence-matrix plan: use rows for OpenAI-hosted, generic self-hosted/customer infrastructure, and each named provider; use columns for control-plane residency, ZDR eligibility, retained session state, execution location, VPC support, file/artifact location, deletion behavior, and provider dependency.
  2. 02Classify each cell only as supported, incompatible, or unspecified from the supplied public documentation.
  3. 03Treat OpenAI-managed agent sessions and execution environments as separate layers; do not infer execution-plane geography from the control-plane residency setting.
  4. 04Count only the bounded launch sample: OpenAI-hosted, generic self-hosted, and nine named provider integrations.
  5. 05Preserve provider-specific retention, residency, and VPC controls as unspecified where the launch documentation does not normalize them.
Next report / 01AI Model Economics Index All research reports