Named agents & workflows

A named agent is a reusable agent profile you set up once and call by name — say, an @researcher that has web tools and a research model, or an @editor with a strict house-style persona. A workflow chains several of those agents (and skills) into a multi-step pipeline that runs on its own.

This guide shows how to create both. For the bigger picture of how MIRA’s agents fit together, see Agents & orchestration.

Create a named agent

Named agents are admin-managed and live server-wide. You build them on the Named Agents admin page.

Go to Admin → Named Agents and add an agent.
Give it a lowercase @handle — this is how you’ll address it (e.g. @researcher). Keep it short and memorable.
Fill in the profile:
- Persona / system prompt — who this agent is and how it should work.
- Tool allowlist — exactly which tools it may use. An @researcher might get web search and fetch; a writing agent might get none. Narrowing the toolset keeps the agent focused and safe.
- Model alias (optional) — pin the agent to a specific model, e.g. a cheaper model for summarising or a stronger one for analysis.
- Budget (optional) — a spend cap for a single run, so a runaway agent stops itself.
Save. The agent is now reusable anywhere on the instance.

Or just ask MIRA to create it

You don’t have to build the profile by hand — describe the agent in chat and MIRA creates it for you. For example:

“Create a MasterResearcher agent that searches multiple sources, pools the findings, and writes a cited report with tables and charts.”

MIRA saves the profile for you (turning a friendly name like MasterResearcher into a handle such as @masterresearcher), picks a sensible tool allowlist, and tells you how to invoke it. The result is a normal named agent — it appears on the Named Agents page, where you can review or fine-tune the persona, tools, model, and budget afterwards.

Use a named agent

Once saved, an agent can be invoked two ways:

By name, in chat. Mention its handle — “ask @researcher to find the latest on X” — and MIRA delegates that turn to the agent’s persona, tools, and model.
Autonomously. MIRA can delegate to a named agent on its own — inside an automation or a proactive task — without you asking each time. Set the agent up once and it becomes part of MIRA’s toolkit.

Running agents show up on the Agents page (the live fleet view), where you can watch per-agent progress, spend, and any artifact files the agent produced.

Chain agents into a workflow

A workflow is a saved pipeline that runs several agents/skills in order, passing outputs forward. You build one on the Workflows admin page.

Each workflow is a set of steps, and each step:

targets a named agent or a skill;
carries a brief — the instruction for that step. The brief can interpolate the run input as {{input}} and any earlier step’s result as {{steps.<id>.output}}, so a later step works on what an earlier one produced;
declares its dependencies — which steps must finish first.

Those dependencies form a DAG (a graph with no cycles). The orchestrator runs it in waves: steps with no outstanding dependencies run in parallel, and each step starts the moment its dependencies finish. So a “gather” step can fan out to three independent research steps at once, then a final “write-up” step waits for all three and combines their outputs.

Make steps resilient or conditional

In the step editor you can mark a step as:

Continue-on-error — if this step fails, MIRA skips the steps that depend on it but lets independent branches finish, rather than failing the whole run.
Conditional (when guard) — the step runs only if an upstream output matches a condition you set. Use this to branch: e.g. only run the “escalate” step when an earlier check reports a problem.
Requires approval (human-in-the-loop) — the run pauses before this step and waits for a person to approve or reject it. The pause is saved to disk, so it survives a restart; approving resumes from exactly where it left off. Use this to gate a step that sends an email or spends real money.

You can also set a per-step budget, the same as for a standalone agent.

Or ask MIRA to build the workflow

As with named agents, you can describe a workflow in chat and MIRA assembles it — “build a workflow that researches a topic with @masterresearcher, then has a writer agent turn the findings into a one-page brief”. MIRA creates the steps (target, brief, and dependencies), saves the workflow, and you can run it straight away or open it on the Workflows page to review and adjust the DAG.

Run a workflow

There are two ways to start a run:

From the UI. On the Workflows page, open the workflow, enter an input, and run it. The page shows run history with live per-step status and each step’s output as it lands.
Conversationally. Just ask MIRA — “run the weekly brief”. It kicks off the workflow, returns immediately, and pings you when the run completes.

Every run is persisted with per-step status and output, so you can open an old run later and see exactly what each step did — useful when a result looks off and you want to find which step went wrong.