Choose what the agent should accomplish
Open Create agent to describe the responsibility in AI Designer, or choose Universal or Specialist and an available role. The conversation can prepare one agent or several connected agents when the work calls for them. Confirm the suggested name or enter your own during setup. Purpose and Working instructions guide its work; the action map defines the operations it can actually perform.
You can build agents for research, content drafting, delivery coordination, financial reviews, lead follow-up and other responsibilities supported by the available actions. A role name is not a restriction to a preset scenario. Each workflow must use actions enabled in Collty and in your connected applications.
| Example | Working route |
|---|---|
| Delivery and financial reporting | Read permitted project evidence, prepare a factual report and send it to the selected recipients. |
| Content and article preparation | Read spreadsheet topics and comments, prepare a work product, attach it to a task, ask for review and revisit the same source record when revisions are requested. |
| Lead follow-up | Read source rows, create a CRM-tagged task and an email draft, wait for approval, send through connected mail, then update the source and CRM from confirmed events. |
| Research and process support | Read authorized documents and operating context, analyze evidence and save recommendations or a deliverable. |
Describe the complete route in AI Designer
Select an agent connected to the project to inspect its role and state. One agent can be used in different project contexts with the relevant project data and permissions.
- 1Select the work area and connections
For project work, connect the agent to the relevant project and enable the required actions in Connections. For a combined company review, open Work areas → Company and choose the saved agent and its sources. A connected MCP service does not automatically grant every operation or access to every file.
- 2State the source and selection rule
Name the actual resource or paste its link. For a spreadsheet, explain the status column and which rows qualify, such as an empty Status. Identify the fields the later actions need: email, topic, comments, dates or a stable row identifier.
- 3Describe results and destinations
Say what to create, where to attach the file, which chat to notify and who receives an email. Distinguish the internal team chat from a shared chat with the client.
- 4State approval, revision and completion conditions
Explain what confirms approval, what triggers rework and which source record should be updated. Sending an email, marking Sent and recording an inbound reply are separate events.
- 5Answer the concrete setup questions
Choose the intended option or give an exact value. Previously clarified settings remain part of the request. Use Build clarified plan when the agreed request is ready for assembly.
- 6Review the proposed map
Inspect the steps, branches and their sources. A statement that the specification is ready is not yet a saved map. Check any remaining resource selections or required fields before Save.
Use the same actions in the visual editor
AI Designer and manual construction produce the same action map. Add a source, action, decision or wait block on the canvas; open its settings and bind the required inputs. Use preceding step outputs for values that are only known during a run.
For a revision loop, keep the source record identity and follow the configured status and feedback condition. A scheduled read alone does not explain which article to revise. A task date must come from a real date field or an explicit fixed date.
For a custom MCP tool, select the exact connection and tool, inspect its input schema and configure required arguments, result bindings and any approval. Do not assume that creating a file also sends mail or that a notification records approval.
Save configuration and activate execution
Save persists the map and its configuration. Start or Pause controls automatic execution in the selected work area. Saving does not silently unpause an agent. AI Designer in Create agent prepares agents; AI Designer inside an action map edits the selected agent’s workflow.
Open Schedule & runs to check the next eligible run and Current work and recent runs. Inspect the actual step results and saved work products after a controlled run. Creation and validation cannot prove that a future resource will remain accessible or that every future row contains complete data.
If assembly or execution needs attention
- A proposed map can remain unchanged when verification fails. Read the concrete issue before retrying the same request.
- Select a real accessible resource and correct required inputs or unavailable actions. A missing email address cannot be inferred from a role or agent name.
- A wait must name an observable approval or revision event. A message asking for approval is not itself approval.
- Check app authorization, project access and AI Credit availability if a previously working operation stops.
- Inspect completed external steps before retrying: an already sent email or a created record is not undone by a later failure.