From one phone to many: planning a multi-device workflow in Swarmbot

Scaling device automation requires more than increasing a device count. Freeze the target list, separate per-device inputs from shared logic, distinguish batches from concurrency and track outcomes individually.

简体中文

Before expanding a working phone workflow, add three things: an explicit target list, per-device inputs and individually verifiable outcomes. Replacing “run once” with “run on everything” is not a rollout plan. The number of selected devices is not the number actually running concurrently.

This guide explains a planning method for Swarmbot users. Relevant implementation was checked on September 10, 2026. This is not a performance benchmark, and it does not prescribe a concurrency setting for every computer, network or phone. No device tasks were executed, and no usable current local activity sample was obtained.

Replace “all devices” with a fixed list

Confirm the intended targets before acting. Groups, filters and the visible wall help you find devices, but execution scope must be explicit. Ask the AI to list its targets for review first.

For Local API integrations, use the device identifiers returned by the current device list, not a card's position on screen. User-assigned labels help humans recognize devices; API identifiers direct requests to the intended target.

Keep a snapshot of the targets, their inputs, the script or skill version and the required starting state. A device appearing later should not automatically join this run. Consult the current client documentation for actual parameters: a planning table is not a universal API request body.

Shared logic does not mean identical inputs

Consider checking whether an application is on an expected screen. The procedure may be reusable, while the app version, login state and starting page differ between phones. Workflows involving files also need distinct outputs to prevent overwrites.

An illustrative planning table—not API parameters—might look like this:

Device alias Input Expected outcome If prerequisites fail
Test A Read-only check Return screen state Human review
Test B Read-only check Return screen state Human review

Do not embed the first device's login details, resource paths or temporary screen state in shared logic. Separate changing inputs, then validate on another representative device to expose assumptions that only held on the original phone.

Total count, batch size and concurrency are different

The total describes how many targets the run covers. Batches divide the rollout into groups that can be reviewed. Concurrency limits work happening at the same time. Submitting many targets does not require starting all of them together.

Swarmbot's current installation-job implementation is a concrete example: it distinguishes requested and effective parallelism, reports queued/running/success/failure counts, and retains per-device progress and errors. Shared resource limits also affect scheduling. This describes installation jobs—not a universal scheduler or parameter set for every feature.

External orchestration systems make similar distinctions. n8n's Loop Over Items emits input items in batches. AWS Step Functions Map supports concurrency limits and cautions against exceeding downstream capacity. These are useful design references, not services required to use Swarmbot. n8n documentation · AWS documentation

Look at the bottleneck before increasing concurrency. Large transfers, image reads and waiting for application screens create different loads. If per-device completion time or errors rise, stop expanding the rollout. Do not bypass limits by launching several overlapping jobs.

Use small groups to test differences, not only speed

A practical progression is:

  1. Validate the complete start, action and result on one authorized device.
  2. Check another authorized device with a different relevant environment.
  3. Try a small group and review each result and resource load.
  4. Pause new targets when a new failure pattern appears.

There is no universal best batch size here. Define acceptance in advance. A read-only check should return a clear state; a phone needing a login or additional permission should go to a human instead of inviting the AI to improvise.

Retain individual outcomes, not just a percentage

Distinguish success, failure, human review and unknown outcome in your business records. These are recommended categories, not a claim that every Swarmbot API returns those exact states.

Retain any task identifier returned by the interface and inspect that original task using its documentation. If device details are paginated, read the relevant pages. Overall progress cannot replace individual results. After a timeout, determine whether the action already happened before retrying it.

Do not repeat successful work just because another device failed. An unknown outcome should not silently become “failed, safe to retry.” For investigation, see our failure-report guide.

An illustrative AI instruction is:

Inspect this target list and each device's prerequisites without executing. After confirmation, work in batches and record individual outcomes. Stop and report targets needing login, permission or outcome reconciliation. Do not add targets or repeat successful work.

A prompt alone does not enforce permissions, concurrency or cancellation. The tools and interfaces used must support those controls. For an external integration, start with the Chinese Local API integration guide and then consult the current client documentation.

参考来源