[Feature Request] Support Webhook/Callback for long-running code execution results

[feifei325]

Is your feature request related to a problem? Please describe.

I am experiencing challenges when running long-duration tasks (e.g., heavy data processing, complex simulations, or long-running Python scripts) that take more than 5-10 minutes to complete.

Currently, the SDK requires a persistent connection to receive execution results through streaming or standard return values. I'm often frustrated when network fluctuations or client-side timeouts cause the connection to drop before the task finishes, making it difficult to retrieve the final output without complex manual polling or state management.

Describe the solution you'd like

I would like E2B to support an asynchronous execution mode with a Webhook/Callback mechanism.

Specifically:

1. When calling sandbox.run_code or exec_cell, I want to be able to provide a callback_url.
2. E2B should accept the task and immediately return a task ID.
3. Once the code execution is finished (whether it succeeds or fails), the E2B backend should send a POST request to the provided callback_url containing the full execution results (stdout, stderr, results, and exit code).

This "fire-and-forget" pattern would significantly improve the reliability of long-running agentic workflows.

Describe alternatives you've considered

1. Polling: Periodically checking the sandbox status or files, but this increases API overhead and is inefficient.
2. Long-lived WebSockets: Keeping a connection open for 15+ minutes is unreliable in many production environments (like serverless functions or behind certain load balancers).
3. Custom Proxy: Building a middle-layer worker to wait for E2B, but this adds unnecessary architectural complexity.

Additional context

This feature is particularly important for Multi-Agent systems where one supervisor agent might trigger multiple long-running sub-tasks across different sandboxes. A native callback support would allow the system to be truly event-driven.

Proposed conceptual API:

sandbox.commands.run(
    "python long_task.py",
    callback_url="https://my-service.com/api/v1/e2b-callback",
    metadata={"job_id": "abc-123"}
)

[linear]

ENG-3421

[riturajFi]

Hi @feifei325 this is super interesting. can i please take a stab at this issue?

[jinngimk-lang]

“Fire-and-forget” is only reliable if the result has an identity that survives the connection dying. Otherwise a webhook just moves the ambiguity from a socket to an HTTP callback.

AgentCI is looking for long-running execution cases like this. The breaker I’d love to preserve is: dispatch task with stable task/attempt ID → drop client connection → completion happens → callback is delivered once or safely retried → caller verifies result digest/terminal state → duplicate callback cannot duplicate downstream effects. Unknown delivery stays explicit instead of being guessed as success.

If you or an external agent build a minimal callback/retry harness, we’ll convert the accepted failure shape into a reusable AgentCI breaker and give durable public credit + backlink to the original issue.

AgentCI: https://github.com/jinngimk-lang/agentci
Breaker/Verifier recognition: https://github.com/jinngimk-lang/agentci/blob/main/COMMUNITY.md

No provider certification claim — this is about proving terminal delivery across a broken client connection.

[devin-ai-integration]

Triage

• type: feature
• priority: low
• complexity: large
• affected_area: API (infra) / command execution delivery
• actionable: yes
• needs_clarification: false
• summary: Request for webhook/callback delivery of long-running code-execution results so callers don't have to hold connections open or poll.

Written by Devin