Skip to content
Empryo

Multi-agent workflows

Connect specialist agents in workflows with separate models, parallel reviews, arbitration and bounded retries. Inspect the graph and follow every run.

Agents & tabs5 minutes to read
Copy & share
On this page8 sections

Choose who does the work, which model each step uses, and what happens after a review. Empryo runs these steps as a graph. A workflow can branch into parallel work, pause for your approval, or return to an earlier step for another attempt.

Start with the jobLink to this section

In the terminal, use /orchestrate and describe what you want:

text
/orchestrate Design a workflow for this project with a planner, a coding agent,
a primary reviewer, an independent security reviewer and an arbiter.
Run the reviews in parallel. Return failed work to the coding agent,
with at most two rounds of fixes. Show me the graph before running it.

This is a request to design the workflow, not a fixed built-in preset. Inspect the resulting graph, model assignments and conditions before starting it. Use models available through your connected providers.

Give each role a clear jobLink to this section

RoleResponsibilityWhat to choose
PlannerPropose the approach and acceptance criteriaA model you trust for design
ExecutorImplement the approved workA coding model
Primary reviewerCheck correctness against the criteriaA separate review context
Independent reviewerInspect the patch without the first reviewer's conclusionsA different model or specialist brief
ArbiterCompare findings and decide whether to accept or request fixesA model you trust to resolve disagreement
text
Plan → Implement → Parallel reviews → Arbitration → Accept
           ↑                              |
           └──────── Fix requested ───────┘

A different model alone does not make a review independent. Give reviewers the task criteria and code to inspect, and decide explicitly which earlier outputs each one receives.

Models, effort and permissionsLink to this section

An agent or judge node can specify its own model and effort. A custom agent can supply reusable instructions, skills and tool restrictions. The router provides defaults for jobs without an explicit assignment.

The orchestration agent designs and coordinates the workflow; it does not itself edit files. Agent steps perform the assigned work. A specialist cannot gain permissions that the session does not allow.

Decisions and limitsLink to this section

  • Agent steps perform work. Judge steps return PASS, FAIL or PARTIAL.
  • Parallel steps start branches. Join steps wait for detached work.
  • Gate steps pause to ask you a question.
  • Conditional edges choose the next step from a result, an answer or saved state.
  • Loops return to previous work. Bound them so a failed review cannot repeat forever.
  • Shell steps can run checks. Inspect commands before running a saved workflow from elsewhere.

A PARTIAL verdict means something remains unverified. Decide how the graph handles it rather than silently treating it as a pass.

Concurrency defaults to 16 agents; the router can remove that configured limit. Provider quotas, memory and workflow limits still apply. An unlimited setting is not a throughput guarantee.

Nodes and edgesLink to this section

An orchestration is a graph of steps (nodes) joined by edges. The node types are agent, judge, parallel, council, advisor, shell, gate, set, link, graph, join and end.

  • An agent node runs a subagent on a task.
  • A judge node reviews the work and ends with PASS, FAIL or PARTIAL.
  • A gate node pauses and asks you a question.
  • A shell node runs a command.

Each edge can have a when condition, so the next step depends on the result: ok, fail, a judge's PASS/FAIL/PARTIAL, your answer at a gate (answer=), or a value stored earlier. An edge that points back to an earlier node makes a loop, and max limits how many times it can repeat.

/orchestrate nightly-sweep            run a saved graph
/orchestrate 5 agents to sweep core   describe a goal: the agent designs the graph, then runs it
/runs                                 watch every run live (terminal UI)

When the agent builds a graph, it can pass it as an object or as JSON text. A graph saved to disk must have its own id.

Follow and reuse a runLink to this section

Use /runs to watch workflows in the terminal. Press v on a run to switch between five views: the map, a timeline, the sequence of steps, the critical path and a plain list. A summary line shows the total time, the critical path, the slowest step and how many retries ran. A view too wide for the terminal falls back to the list.

With agentFeatures.subagentTabs on, each agent step opens as its own child tab under the chat that started the run, so you can read or steer it like any chat.

On desktop, use the Orchestration area of the workbench. The orchestration tool supports validation, saving, running, watching, resuming and stopping workflows.

A saved graph can be run again with /orchestrate <id>. Review it after changing model assignments or project requirements. Start with a small task before scaling out.

When something goes wrongLink to this section

Runs survive a quit or a crash. When Empryo starts again, a run that was working picks up where it stopped, and a notice says how many resumed. Agent steps continue in the same tab, and gates, parallel branches, nested graphs and detached agents carry on from their last state.

A provider problem pauses the run instead of failing it. When a step's model needs you to sign in again or its plan has run out, the run pauses and waits for you, without spending its review loops on an error that is not in the code. A short outage is retried after 1, 2 and 4 minutes; after that, the run waits for you to retry it. Finished parallel branches are kept when it retries.

Restart from a chosen point. Ask the agent to restart a failed, stuck or interrupted run. It re-enters at the step where the run stopped, or at a step you name, and can clear the loop counters, raise the loop limit or set saved values first. Earlier results stand.

Gates pass on their own when the tab approves everything. In a tab on YOLO, approve-all or full approvals, approval gates do not stop to ask.

Pick the simplest tool that fitsLink to this section

Use parallel agents for independent pieces of one task. Use a goal loop for repeated implementation and fresh review. Use a workflow when you need explicit roles, branching decisions or approval points. For a second opinion without an execution pipeline, use an advisor or council.