Sandboxes

Each trial runs in its own cloud sandbox. Use -e to request E2B, Daytona, or Modal; the default is Daytona.

evolve run \
  -d deepswe@1.1 \
  -i anko-default-function-arguments \
  -a codex \
  -m gpt-6-luna \
  -e modal \
  --max-trial-spend 1 \
  --max-retries 0 \
  --watch

Task requirements come first

Evolve checks the requested provider against the task’s image, resources, network policy, Compose services, and GPUs. A fallback is recorded as sandbox_provider_degrade on the trial. The requested provider on the job and the actual provider on a trial can differ. If no provider can satisfy the task, Evolve rejects the configuration.

evolve dataset show deepswe@1.1

Inspect the per-task provider verdicts before starting a larger run. A compatible declaration does not guarantee immediate provider capacity.

Supported configurations

Capability	E2B	Daytona	Modal
Single-container tasks	Yes	Yes	Yes
Docker Compose	Yes	Yes	No
GPUs	No	Subject to live quota and GPU type	Yes
Compose with GPUs	No	No	No
Compose with `no-network`	No	No	No

Network allowlist kinds and resource ceilings also differ. The live meta() document publishes these constraints. Daytona disk and GPU limits depend on the platform’s live provider quota.

A temporary capacity shortage can put a trial back in the queue even with --max-retries 0. Capacity waits are bounded and separate from ordinary retry attempts.

Your own provider account

Add --own-infra to run trial sandboxes in your E2B, Daytona, or Modal account. The CLI reads the provider’s credentials from your shell, or uses stored secrets when a variable is absent.

This option requires server support. If the server returns feature_unsupported, omit --own-infra to use Evolve’s account.

Provider	Credentials
E2B	`E2B_API_KEY`
Daytona	`DAYTONA_API_KEY`
Modal	`MODAL_TOKEN_ID` and `MODAL_TOKEN_SECRET`, or the active profile in `~/.modal.toml`

Daytona also accepts DAYTONA_API_URL on a daytona.io host and DAYTONA_TARGET of us (default) or eu. Custom regions are not supported.

export MODAL_TOKEN_ID=ak-...
export MODAL_TOKEN_SECRET=as-...

evolve run \
  -d deepswe@1.1 \
  -i anko-default-function-arguments \
  -a codex \
  -m gpt-6-luna \
  -e modal \
  --own-infra \
  --max-trial-spend 1 \
  --max-retries 0 \
  --watch

Credentials from your machine are saved as secrets on first use. Either delivery mode works; the keys never enter a sandbox. The CLI reports whether each key came from your machine or stored secrets.

Limit	Own-account behavior
Images	Public images only. Public datasets qualify; private dataset images are held in Evolve’s private registry.
Task types	No Compose tasks or Daytona GPU tasks.
Provider selection	No fallback. An incompatible task is rejected with `provider_unsupported`.
Concurrency	The job’s `-n` limit, your organization’s shared concurrent-trial limit (default 16), and your provider’s limits apply. Evolve’s per-provider sandbox ceilings and GPU cap do not.
Resume and retry	Only the creator can start them because they use that person’s keys.

The job records sandbox_account: own. Its regrades retain that attribution, but analysis, checks, and regrade sandboxes run on Evolve’s account. The SDK equivalent is sandbox_credentials on jobs().start(). You can access your provider’s sandboxes, including hidden tests and reward files. This access affects the trust others can place in the scores.

Replace or remove credentials

Deleting a credential makes the remaining trials of jobs using it fail. Those jobs do not switch to Evolve’s account.If your shell holds a different value from the stored secret, the request is refused. Unset the variable to use the stored key. To rotate it, wait for your own-account jobs to finish, delete the stored key with evolve secrets delete, then run with the new value.

Sandbox cleanup and reusable images

Evolve ends each sandbox when its trial ends or is stopped. If normal cleanup fails, the provider lifetime applies: the trial time budget plus a margin on E2B and Modal, or Daytona’s auto-stop followed by deletion.Reusable images remain in your account:

Provider	Retained resources
E2B	Templates named after the task image, with an `-o<id>` suffix
Modal	The `evolve-sandbox` app and its images
Daytona	Cached images

You can delete these in your provider console. A later job rebuilds what it needs.

Images and startup

Dataset publication prepares task images. A provider may also need to prepare its own cached image or snapshot before the first trial. That first start can take longer than later trials using the cache. Resources come from the task’s task.toml. See task configuration for defaults and how to request CPUs, memory, disk, and GPUs.

GPU costs

GPU compute is shown as an estimate, separate from model spend:

Location	Field
Trial	`gpu_cost`
Job totals	`stats.gpu_cost_usd`
CLI trial view	`gpu compute (est.)`

An unavailable estimate includes unpriced_reason. Do not add this estimate to the metered model cost and present it as the same charge.

Analysis and checks

These operations also use sandboxes. Choose their provider with -e on evolve analyze or evolve check, or --analyze-provider for analysis attached to a run. Their default is Daytona, but server defaults can differ. Read the resolved provider through the SDK defaults.

Start here

Run evaluations

Task format

Validate your work

Read results

Work together

Task requirements come first

Supported configurations

Your own provider account

Images and startup

GPU costs

Analysis and checks