Choose a ready solution and describe the task

Before you start

Browsing is public; running, durable history or installation can require sign-in.

Click path

1. Open IAM.Market, choose EN, then Solve now for a one-time task.
2. Click Catalog to see ready solutions.
3. Select the intended card and click its Run control.
4. Describe the goal, constraints, available inputs and desired output.
5. Use Attach for approved documents, then Send request.
6. Complete sign-in if requested and review the result, sources and limitations.

Enlarge screenshot

1 2 3 4 5 6 7

1. Запустить
2. Запустить
3. Запустить
4. Запустить
5. Запустить
6. Запустить
7. Каталог

Verify and recover

Verify a result for your task. A grocery recommendation is different from retailer cart transfer or checkout. Inspect History before retrying a failed run. Commercial access and integration permissions can be prerequisites.

Subtasks and result verification

Open My agents, select an installed agent, and describe a complex task with its goal, constraints and acceptance criteria. The agent creates an Execution plan with up to eight stages, dependencies and individual acceptance criteria. Simple questions remain ordinary conversations.

Up to three independent read-only stages can run together. Stages that change data run one at a time. Each stage receives only permitted tools available in its actual runtime; an unavailable integration cannot be replaced by a claim that an action succeeded. A separate reviewer checks each result, and tool-backed stages require actual host observations. The overall result is then checked against the original task.

Enlarge screenshot

1 2

1. План выполнения
2. Результат проверен

This educational example shows synthetic checks on staging. It illustrates stage states, not customer data processing.

Duration, stopping and continuing

Each stage can run for up to 20 minutes including its review. The entire plan has a 60-minute work budget; the chat allows up to 65 minutes for completion and response delivery. Each planning attempt and individual reviewer request have a two-minute limit. A malformed reviewer response allows one repeat review of the same result without rerunning actions; a negative verdict is not automatically retried. An invalid plan has up to three correction attempts; no stage is dispatched until the plan is validated. Hosted agents allow up to 256 model-and-tool iterations per stage. Tenant/token quotas, model availability and repeated-failure protection still apply. Reinstalling the agent is unnecessary. The specialized grocery workflow retains its own execution deadlines.

Click Stop execution in the plan to stop work. Accepted results are persisted immediately, even while another stage is still running. Closing a tab does not cancel the task: return through History to inspect its saved state.

Enlarge screenshot

1

1. Остановить выполнение

If a plan finishes partially, inspect completed and unverified stages. When safe, click Continue unfinished stages: the agent reuses accepted results and runs only the remaining work. A read-only stage rejected by its reviewer gets one repair attempt. If planning fails validation, clarify the goal, available tools and acceptance criteria in a new message. Automatic continuation is unavailable when an interrupted action may have changed external data: inspect the actual service state first, then provide a concrete next instruction. An error or cancellation does not confirm that requested changes were saved.

Enlarge screenshot

1

1. Продолжить незавершённые этапы

If every stage is marked Result verified but final acceptance did not finish, click Retry final review. The reviewer reassesses the saved accepted results; workers and their actions are not run again.

Enlarge screenshot

1

1. Повторить проверку итога