Skip to content
OpenCorvus

How Missions run long-horizon work

How OpenCorvus coordinates fixed-squad Tasks, typed handoffs, evidence, and reopening across a long-horizon outcome.

A Mission turns one long-horizon outcome into a chain of owned, reviewable Tasks.

Mission coordinates the outcome. It is not a larger Agent. Each child Task owns one exact Expert Squad revision and one selected workflow, while dependencies, messages, Artifacts, evidence, and lifecycle decisions all stay visible.

Coordination stays separate from execution

Mission owns the outcome and the dependency graph. The work, the capability projection, and the evidence stay with their real owners.

OwnerOwnsDetail
MissionThe outcome and its dependenciesCoordinates outcomes across child Tasks and records their dependency graph.
TaskOne owned deliveryOne project-scoped delivery, one fixed Expert Squad revision, the selected workflow, the related Sessions, and its lifecycle decisions.
Expert SquadRoles and capability projectionNamed agents, instructions, Skills, tools, MCP access, the selector, and declared workflows.
Artifacts and evidenceExact, reviewable handoffTyped Artifacts carry provenance and exact references. Host observations record file and command facts independently of any agent’s summary.

How it runs

  1. Define the outcome. Record the result wanted, the task boundary, and the acceptance conditions.
  2. Split into owned Tasks. Split only when a delivery can be independently owned, accepted, or depended on. Splitting for its own sake produces coordination overhead with no owner.
  3. Hold the squad identity set. Mission records which Expert Squad IDs are available when it starts. Capabilities installed later do not silently widen that set.
  4. Freeze the revision per Task. When a child Task is created, an admitted squad ID resolves and pins to one exact package revision plus its selected workflow, for the life of that Task.
  5. Run in-task collaboration. Named roles stream messages and tool work, then hand off typed Artifact references with visible evidence.
  6. Read the evidence and decide. Mission reads the outputs of terminal child Tasks and continues through explicit acceptance or a dependent-Task path.

Why long-horizon work holds together

  • Persistent organization. Mission and Task state are persisted. Versioned outcomes and acceptance contracts preserve what is being delivered and judged.
  • Fixed Task authority. One exact squad revision and its selected workflow stay fixed for the Task’s lifetime. A Task cannot silently switch squads mid-life.
  • Handoff with evidence. Typed Artifact references, provenance, complete reads, and host observations let later work inspect exactly what a decision rested on.
  • Terminal is not final. A Task that reached completed, failed, or cancelled reopens when real new input arrives — sending a message is the resume. Reopening starts a fresh execution occurrence rather than replaying the old one, and the prior history stays readable. There are no separate retry or replan controls to find.

The real boundary

Long-horizon is not unlimited execution, and it is not permanent unattended operation.

Unattended work continues only while the local or hosted OpenCorvus runtime is online. Output quality depends on the selected model, the reachable sources, the installed capabilities, and the available evidence. External effects remain bound by permission rules. Mission coordination does not erase Task or agent ownership.