# Build an Agent you can test, reuse, and hand over. Reviewed: 2026-10-18 Canonical: https://clankercloud.ai/learn/ai-agent-factory An AI agent factory is a repeatable process for creating and maintaining useful Agents. In Clanker Workspace, define a role, add its instructions, reference knowledge, Skills, and selected Tools, test its saved work, then reuse or share the reviewed Agent. The factory is the working process around the Agent; it does not mean training a new foundation model. ## What you are working toward A repeatable path from a task brief to a tested, maintained Agent and a reviewable handoff. ## Start with - A repeated job with an owner who understands the expected result - A normal example, a difficult example, and explicit action boundaries - A Workspace Project for inputs, saved outputs, and review notes ## Make the factory produce a useful role Start with a job that has recognizable inputs and an output someone can check. A proposal reviewer, reporting specialist, or intake analyst is easier to assess than an Agent asked to run an entire company. Name the owner who will decide whether the work meets the brief. The unit you are creating is a reusable method with selected capabilities. Keep a changing client name, this week’s reporting period, and current source files in their Project. Keep the role’s evidence rules and recurring review steps in the Agent. This makes reuse understandable when the next task arrives. ## Assemble the smallest complete Agent Create or customize an Agent in My Agents. Give it a clear job, instructions, and stable reference knowledge. Select a Skill for a reusable method and a Tool for an operation the job needs. Record why each component is present so a later reviewer can distinguish a missing input from a missing operation. Personality, voice, and appearance can make a specialist recognizable, but they do not establish competence. A fictional professional profile is not a credential. Judge the Agent by its work against the supplied evidence. Account connections and tool access remain separate configuration choices; a sentence in the brief does not grant them. ## Run a small acceptance process Use one fictional or approved packet with a known answer. Ask for a saved deliverable, then open that file in Project Files or Preview. Check the material claims, calculations, omissions, and requested format. Record what you observed instead of accepting a confident completion message as proof. Run a second case with a missing fact or conflicting source. A useful Agent should preserve uncertainty and identify what it needs. Manual review is the baseline for this process. If saved evaluations are enabled for your Workspace, inspect their available checks and evidence; do not assume that every runtime or task is covered. ## Maintain behavior with a reason for each change When a test fails, identify the specific failure before adding more instructions or tools. A wrong date may come from a stale source, an unclear precedence rule, or a missing review step. Write a narrow correction and repeat the unchanged case so you can tell whether the change addressed that failure. Save the improved Agent and use version history when you need to restore an earlier setup. Keep a short review note with the problem, changed rule, tested inputs, and observed outcome. An instruction that helps one example can damage another, so retain the earlier cases as the role develops. ## Choose a distribution route deliberately A personal Agent can remain private. For another person to add a copy, review the private sharing route and included dependencies. Marketplace publication is a separate reviewed version. Inspect the current sharing package and recipient flow before assuming the receiver has the same configuration or access. Do not treat copying an Agent as transferring connected accounts, secrets, or every source file. The recipient needs appropriate inputs and separately configured access. Give them a small test packet and the acceptance checklist. Their successful first run is more useful than a promise that the copy will behave identically in every account. ## Expand only after the original job works A reviewed role may later support a private App, a specialist handoff, or configured recurring work. Test the new workflow separately. A good draft report does not prove that an App preserves edits or that a Schedule handles an unavailable source. Hosted compute and inference have usage costs that need a deliberate budget. Choose the execution environment for the actual job. A native Agent, an external assistant using MCP, a BYOA harness, and a Sandbox API runtime are different surfaces. Use the environment guide to choose one, then keep sources, saved results, and the next task’s handoff together in the Workspace. ## Try this prompt Copy into a Workspace Session. Opening Workspace does not send it automatically. ```text Help me use Clanker Workspace as an agent factory for one repeatable job. Ask for the input packet, desired saved output, reviewer, and actions that need a separate decision. Draft an Agent specification, propose only the necessary knowledge, Skills, and Tools, and give me one normal test and one missing-input test. Start with manual review of saved files. Do not publish, share, connect accounts, or configure recurring work from this request. ``` ## Before using the result - The Agent has one defined job and a named reviewer. - Each selected component supports a specific task step. - A normal case and a difficult case have recorded observations. - The actual saved deliverable opens and matches its sources. - Sharing, publication, and recurring work are separate reviewed choices. ## Common questions ### What does AI agent factory mean in Clanker Cloud? It means a repeatable process for designing, configuring, testing, maintaining, and distributing custom Agents in Workspace. It does not mean training a foundation model or guaranteeing that every configured Agent will perform correctly. ### Do I need automated evaluations to start? No. Start with known examples and manual review of saved outputs. Saved evaluation and learning features depend on Workspace availability and supported runtime behavior; inspect the available controls before relying on them. ### Does creating an Agent publish it? No. You can test it privately. A private copy and a reviewed Marketplace publication are separate distribution choices with their own included content and access requirements. ## Sources - [Clanker: create and test a custom Agent](https://clankercloud.ai/learn/build-a-custom-agent) - [Clanker: knowledge, Skills, Tools, and access](https://clankercloud.ai/learn/agent-knowledge-skills-tools) - [Clanker Workspace: current product capabilities and availability](https://clankercloud.ai/faq) - [Clanker: current pricing and usage](https://clankercloud.ai/pricing) ## Related - [Give your agents a place to do lasting work.](https://clankercloud.ai/learn/ai-agent-workspace) - [Choose an environment by what the agent must reach.](https://clankercloud.ai/learn/choose-ai-agent-environment) - [Test the work before you reuse the Agent.](https://clankercloud.ai/learn/test-ai-agent-before-reuse) - [Turn a correction into a better repeatable method.](https://clankercloud.ai/learn/improve-agent-from-corrections) - [Share a reviewed method, then test the recipient’s copy.](https://clankercloud.ai/learn/share-custom-agent-with-team) - [Write the Agent brief before adding capabilities.](https://clankercloud.ai/templates/agent-specification) - [Give the Agent a test with an answer you can check.](https://clankercloud.ai/templates/agent-acceptance-test) - [A small packet the next agent can verify.](https://clankercloud.ai/templates/agent-handoff-pack) [Open Workspace](https://workspace.clankercloud.ai) [Current pricing and usage](https://clankercloud.ai/pricing)