What Matters
An agent can use a standard tool protocol and still be expensive to move. Its session state, approvals, identity, recovery rules and operational history may belong to different systems. DN’s Runtime Portability Index maps those boundaries and provides a migration worksheet. This first edition documents architecture; it does not claim measured migration times or rank vendors by an invented score.
DN Evidence Block
- Last verified: 19 October 2026. The 2027 title is an editorial planning horizon, not a future observation date.
- Scope: OpenAI Agents API and SDK; Amazon Bedrock Managed Agents powered by OpenAI; Google ADK deployment documentation; Microsoft Foundry Agent Service; LangGraph persistence.
- DN contribution: A six-boundary migration framework, a documentation index and a local evidence worksheet.
- Limits: No cloud accounts were deployed or migrated. No comparable latency, migration-cost or task-completion results were measured.
- Prepared by: Decentralised News research desk. No independent reviewer is claimed. Submit a correction.
The model can move while the agent stays behind
Imagine a finance assistant that reads invoices, requests approval and writes a reconciliation record. Replacing its model endpoint may take a small adapter. Moving its partially completed work is a different task: the new environment needs the original principal, pending approvals, tool results, artifact references and a reliable account of which external actions already happened.
A transcript is not necessarily a resumable execution state. A container image does not include every cloud identity or managed data service it calls. A valid credential in the new host does not establish that the old spending limit is enforced. Those are separate migration questions, and each needs evidence.
DN’s definition: practical portability means reproducing the workload's accepted behavior and authority limits in the destination, with a documented transition and rollback. It does not require identical internal implementation. A migration is incomplete when it changes the business acceptance criteria without making that change explicit.
Six boundaries that determine migration burden
| Boundary | What must be accounted for | Useful evidence |
|---|---|---|
| Agent definition | Instructions, tool schemas, model settings and dependency versions | A pinned source bundle and target-compatible configuration |
| State and memory | Session records, checkpoints, stores and artifact references | An export/import map and a resumed task checked against an independent record |
| Tools and identity | Endpoints, credential audiences, scopes and principal mapping | Least-privilege credentials reissued for the destination; no copied secrets |
| Approval and policy | Who approves which effect, limits and revocation | The approved descriptor still matches the dispatched action |
| Recovery and evidence | Retries, idempotency, receipts and unresolved outcomes | A lost-response rehearsal with no unapproved duplicate effect |
| Operations and economics | Regions, logs, billing, cleanup and human work | A cost ledger, observability plan and tested rollback |
What the current documentation actually establishes
| Platform path | Documented observation | DN migration question | Evidence |
|---|---|---|---|
| OpenAI Agents API / SDK | Service-managed sessions and an application-controlled SDK are different choices. | Inventory which state and execution responsibilities your chosen path owns. | Official architecture comparison |
| Amazon Bedrock Managed Agents powered by OpenAI | AWS labels the service public preview; conversation and execution-environment files have separate lifecycles. | Treat IAM, environment storage and preview limitations as explicit migration items. | AWS preview documentation |
| Google managed runtime with ADK | Current ADK documentation describes deployment of agent code and dependencies to a managed runtime. | Verify the exact service/API version; deployment portability alone does not establish state transfer. | Google ADK deployment guide |
| Microsoft Foundry Agent Service | Microsoft distinguishes configuration-led prompt agents from hosted agents that carry application code. | Specify which agent type, identity and attached services are being moved. | Foundry documentation |
| Self-hosted LangGraph | Checkpointers and stores serve different persistence scopes. | Record the chosen backends, checkpoint compatibility and recovery rules. | LangGraph persistence guide |
Google's older Agent Engine documentation paths currently redirect to newer Agent Platform/Agent Runtime material. This is a reason to pin the exact source and API used in a migration record, not evidence that every historical endpoint or workload has already migrated.
Who should investigate which path?
A managed harness can fit teams seeking less orchestration infrastructure to operate. Application-owned frameworks can fit teams that need explicit control of execution and state. Hosted custom code can sit between those choices. These are architecture fits, not DN performance findings. Avoid choosing from an umbrella product name alone: the model, agent type, storage service and sandbox can change the trade-off.
All routes have costs. Managed services can bill inference, tools, compute, storage or networking; self-hosting adds operations, patching and capacity responsibilities. Access depends on the actual account, region and supported features. None of the documentation observations here establishes custody suitability, production security or eligibility for a financial use case.
DN Migration Burden Worksheet
Record your own evidence for one pinned workflow. These are your declarations, not DN findings, a safety score or a scan. Unknown is the default. Do not enter credentials or private task content.
All boundaries start unknown. The surrounding table explains what evidence to collect; this worksheet does not inspect a provider.
Runs locally with no account, wallet, model call or storage of your answers. A downloaded worksheet contains your selections. Normal DN site analytics are separate.
Run a reversible migration rehearsal
- Freeze one bounded workload, its model/tool versions and acceptance rules.
- Save a redacted source record and identify every external effect already committed.
- Map identity, state and approvals to the target; leave unexplained fields unknown.
- Replay a read-only or synthetic task before permitting any external write.
- Interrupt a permitted test after dispatch but before acknowledgment. Reconcile independently before retrying.
- Compare accepted outcomes, total cost and human recovery work. Keep an explicit rollback path.
The rehearsal is a proposed method. A worksheet selection does not prove that these steps were performed. A successful synthetic task does not establish production equivalence across all workflows.
Price the exit while choosing the entry
Migration burden is an economic dependency: stored history, hard-to-export artifacts and specialized approval logic can create future work even when today's token price is low. Use the DN Agent Cost Stack to keep implementation and human recovery costs visible. Use the delegation integrity tests to define authority checks that must survive a move.
DN Alpha Thesis
The useful portability asset is an evidence-backed exit plan: a clear account of what can move, what must be rebuilt and what has not been tested. Providers that make those boundaries inspectable can reduce evaluation friction. That is DN’s research thesis, not a forecast that open interfaces remove lock-in.
Methodology, limitations and change log
Version 0.1, 19 October 2026. This index classifies documentation and proposes a workload-specific assessment. It has no vendor league table or composite score. Source pages are living documents; preview support and product naming can change. A later release can add timed, version-pinned migrations with independent outcome checks. Missing observations will remain null rather than zero.
For the broader architecture comparison, read the Managed Agent Runtime Index. To propose a reproducible migration study, contact DN with a bounded workload and public evidence.
Primary sources
- OpenAI: agents architecture choices
- AWS: Bedrock Managed Agents public preview
- Google ADK: deploy to Agent Runtime
- Microsoft: Foundry Agent Service
- LangGraph: persistence
Frequently asked questions
Does MCP support make an entire agent portable?
No. A tool interface does not transfer session state, credentials, approval semantics, recovery behavior or billing.
Does this index rank the safest runtime?
No. Version 0.1 is a documented-boundary index and planning method, not an executed migration benchmark or security rating.
Can self-hosted frameworks eliminate lock-in?
They can expose more implementation choices, but dependencies, checkpoint formats, model behavior and operational responsibilities still need evaluation.
Disclosure: Educational research, not investment advice or a security certification. No affiliate or sponsored placement appears in this edition. DN may discuss scoped research services; payment cannot purchase a finding. Corrections should identify the affected claim, observation date and supporting source.
