Which Coding Agent Stack Recovers Without Duplicate Work?
Shortlist orchestration, execution, and workspace layers separately. Compare complete stacks with three interruption trials and require exactly one review record for the intended artifact.
Compare the responsibilities each option actually covers
Use OpenAI’s Agents API if you want a managed Codex harness, Agents SDK if your application should control handoffs, and Responses API if you want to build the loop against model responses. Microsoft’s AutoGen repository is archived and points readers toward Microsoft Agent Framework. Treat them as separate maintenance choices and record who will own fixes for each. Cloudflare’s OpenCode sandbox example packages repository edits and diff retrieval. Claude Managed Agents offers hosted or self-hosted tool execution while Claude still chooses tool calls. NanoClaw exposes a scheduler and per-group containers in its repo. Those eight options span different layers. Compare complete deployments by assigning an owner for work dispatch, artifact storage, and review publication. In an illustrative trial, measure the delay from worker interruption until the reviewer can retrieve the original artifact. A replacement worker starting quickly is only one component of that delay.
Agents | OpenAI APIForce the review receipt and artifact to disagree
Use Cloudflare’s OpenCode example as a concrete producer for the test shape. Its output is a repository diff, which gives your reviewer something independently inspectable. Use a tiny fixture change and assign the normalized diff a digest. In a disposable environment, pause execution once before artifact storage, once after storage but before review publication, and once after review publication but before acknowledging the job. Replay the same job ID and read the artifact from the reviewer’s path. Require the expected digest, one review record, and a measured recovery duration for each cut. The last cut exposes a retry that republishes review even when the previous publication succeeded. The middle cut catches a worker that regenerates a plausible but different patch. Keep the expected digest fixed for the fixture so either failure appears in the result.
Run OpenCode in a sandbox · Cloudflare Sandboxes docsDecide who owns recovery state
OpenAI’s Agents API runs a managed Codex harness and is positioned for long-running work with provider-managed progress. The OpenAI Agents SDK runs inside your application and gives you control of the agent loop, tools, and handoffs. Prefer the managed API for an initial test when you want the provider to manage agent progress. Choose the SDK for the pilot when your application must control what happens before review begins.
For both options, define the record you must recover after a restart. Include the job ID, input revision, artifact reference, validation result, and approval state. Even with managed progress, verify whether those records are available after the interruption you expect. Keep any required review state in an application record if the platform cannot demonstrate that it persists.
Agents | OpenAI APIInterrupt the handoff with a known artifact
In an illustrative pilot, ask a coding agent to correct three broken relative links in a sample documentation repository. Give it a base revision and an approved list of documentation files. Require a diff and a passing link check. Interrupt the orchestrator once the diff has been recorded and before review begins. Resume the same job ID and confirm that review receives the same diff and validation result exactly once.
Use the managed Codex harness and the application-owned Agents SDK as reference implementations for the test. Run the same interruption against LangGraph and CrewAI if they are on your shortlist. In each candidate’s documentation for the edition you would deploy, verify checkpoint storage, retry behavior, and approval handling. Count a repeated edit, missing artifact, or duplicate review request as a failure.
Agents | OpenAI APIAccount for maintenance on Microsoft’s path
For an existing AutoGen workflow, put Microsoft Agent Framework on the shortlist and treat the old installation as a migration baseline. Microsoft’s AutoGen repository says the project was archived. Inventory the actual workflow before attempting a port. Record which component selects a role, where tool permissions are enforced, and what data the reviewer gets when an author hands off a change.
Run a small existing job through both implementations and deliberately reject the first review. Compare whether the same artifact is available after rejection, whether the author receives the reviewer’s reason, and whether a corrected artifact gets a fresh approval. If the replacement changes the handoff, estimate the application work needed to restore that behavior before deciding when to migrate.
GitHub - microsoft/autogen: A programming framework for agentic AI · GitHubSeparate hosting from handoff logic
Cloudflare documents running OpenCode in a Linux sandbox to edit a repository and read the changes as a diff. This path is a useful baseline when the workflow needs one bounded editing job followed by human review. Try it with the same sample documentation task, return the diff, and measure which requirements still need application code, including job submission, durable artifact storage, review status, and recovery after a worker interruption.
If one job and a reviewer cover the workflow, compare that smaller implementation against the extra state you would operate with multiple agents. If a separate reviewer agent is required, send it the immutable diff and validation result. Fail the pilot if it receives an artifact from the wrong revision or can change the recorded approval.
Run OpenCode in a sandbox · Cloudflare Sandboxes docs