# An agent runtime with a task you can inspect. Reviewed: 2026-10-20 Canonical: https://clankercloud.ai/learn/ai-agent-runtime Clanker Cloud provides a hosted work environment that external agents can operate through authorized tools. A managed Session records the task; an execution tool can start a bounded compute job; saved files preserve its results. ChatGPT, Claude, Hermes, and OpenClaw keep their own agent loop unless you separately configure a supported hosted Agent. ## What you are working toward One bounded run with a visible task, an inspected job receipt, and verified saved outputs. ## Start with - A specific task and a known expected output - Execution access only if the task needs compute - A plan for checking results and stopping failed work ## Know which layer is running | Layer | Responsibility | Evidence to inspect | | --- | --- | --- | | External agent client | Reasoning, conversation, and decisions about tool calls | The client’s actual calls and results | | Managed Workspace Session | Visible task identity, public progress, and lifecycle | Session/run identifiers and final recorded outcome | | Workspace compute job | A submitted command or script with its limits and expected outputs | Job status, errors, validation, and saving receipt | | Saved Project files | Durable outputs a later task can open | File identity, revision, content, and Project | ## Start a managed task before changing work Read the current instructions and discover exact tool schemas. Before a change or execution, start a managed Session with a stable task key and retain its returned identity. Pass sessionId to changing calls. Send concise public progress heartbeats while work continues, including during job waits. The Session gives the owner a place to watch and take over the external task. It is not itself a shell, a model process, or proof that any compute job ran. An expired, completed, revoked, or taken-over task cannot simply be reopened by repeating its start call. Stop and resolve the actual state before doing more work. Sources: - [Clanker: current agent instructions and access contracts](https://workspace.clankercloud.ai/for-agents.md) ## Treat execution and saving as separate stages For a small CSV-processing task, specify approved inputs, the transformation, a time allowance, and expected output files. Preparation can consume part of a job’s time allowance before user code starts. Inspect the original job rather than submitting a replacement because a request timed out. A successful command exit does not establish a saved deliverable. Check validation and saving status, then read the committed output files. If setup fails before execution, report that user code did not run. If a result was not saved, do not present a temporary working file as a Project artifact. Discover current input-selection and output-validation arguments instead of assuming defaults. ## Recover by identity, not by repeating effects Retain task, job, file, and action identifiers. Reconcile an uncertain outcome using the existing operation and its stable key. Reread the current revision before editing an existing file. When a stale revision is refused, select the current source rather than weakening the check. Read bounded output pages when a Session has more saved files than fit in the first response. Follow the returned cursor and compare file identities and revisions; the first page is not necessarily the entire result. Finish with a public final message, verify it, and complete the Session with an honest succeeded or failed status. ## Choose a hosted Agent or a standalone runtime deliberately Connecting a client through MCP does not install that client inside Clanker Cloud or move its model inference. Supported bring-your-own-agent environments are a separate deployment choice with their own configuration, task execution, saved-state behavior, and costs. Native Workspace Agents are another way to run a reusable method. The standalone Sandbox API has its own runtime identity, credentials, capabilities, and expiry. A Sandbox ID is not a Workspace Session ID. Saved Project files can outlast active compute, but unsaved process memory is not a durable handoff. Use the appropriate contract for the environment you actually selected. ## Try this prompt Paste this prompt into the external assistant connected to your Workspace. The connection guide does not launch or send a task automatically. ```text Help me plan one bounded Clanker Cloud Workspace runtime task using a fictional CSV I provide. Discover the current command and job schemas first. Before execution, show me the approved input paths, exact transformation, timeout, expected output files, and usage implications. After I authorize the run, start a managed Session, submit once with a stable action key, inspect the original job, read back saved outputs, and complete honestly. Stop on takeover or revoked access. Do not install packages, publish, or schedule work. ``` ## Common questions ### Does connecting ChatGPT host ChatGPT inside Clanker Cloud? No. The connected client continues its own conversation and model calls. Clanker provides authorized Workspace tools and selected execution environments. A supported hosted Agent or BYOA deployment is a separate setup. ### Is an agent runtime always running? Do not assume so. Sessions, bounded compute jobs, hosted Agent deployments, standalone sandboxes, and recurring workers have different lifetimes. Check the selected environment’s limits and preserve important work as saved files. ### When is a run complete? After the requested work has actually run, required outputs have been saved and inspected, and the managed Session has a verified final message and terminal status. A queued job or a zero exit code alone is incomplete evidence. ## Sources - [Clanker: current agent instructions and access contracts](https://workspace.clankercloud.ai/for-agents.md) - [Clanker: current native tool schemas and availability](https://workspace.clankercloud.ai/workspace-tools.json) - [Clanker: choose an agent hosting route](https://clankercloud.ai/learn/run-agents-in-cloud) - [Clanker: separate standalone Sandbox API lifetime and recovery](https://clankercloud.ai/api-docs-agent.md) ## Related - [A cloud for agents, with the work kept together.](https://clankercloud.ai/learn/agent-cloud) - [Choose an environment by what the agent must reach.](https://clankercloud.ai/learn/choose-ai-agent-environment) - [Run the task in the cloud. Keep the result for later.](https://clankercloud.ai/learn/run-agents-in-cloud) - [/workspace-agents](https://clankercloud.ai/workspace-agents) [Open Workspace](https://workspace.clankercloud.ai) [Current pricing and usage](https://clankercloud.ai/pricing)