Coordinate agents with a plan you can check first.
AEGIS Orchestrator is designed to turn an objective into a stored, validated plan, and to delegate each step with only the authority it needs. Models may help draft a plan. They do not get the final say on whether it is valid.
Task decomposition
The Planner turns one objective into a versioned plan of steps. A plan is data, so it can be validated before any step runs.
Plan validation
Bad plans fail before they cost anything.
Plan v1: invoice review, 6 steps
Plan valid
Schema, agent versions, dependencies, Mandates and budget all check out, so the plan is stored and scheduled.
Simulated. Error text shows the intended style.
Two kinds of plan
Autonomy expands as evidence allows.
Static
A developer defines the graph. The run fills in the data.
Use when the process is known and repeatable, such as month-end reconciliation.
Bounded planner
A planning step chooses from allowed templates, and the output is validated.
Use when the approach varies but its limits are clear, such as support triage.
Dynamic replanning
The plan can change mid-run. Each new version is validated and cannot widen authority.
Use when evaluations show that it is safe for the task.
Example workflow
What happens to an objective.
Receive
An objective arrives with a root Mandate and a principal.
Compile
A definition or planner produces a draft graph of steps.
Validate
Schema, published agents, dependencies, budget and Mandates are checked.
Schedule
Steps with no unmet dependencies become ready, and workers lease them.
Join or replan
Results flow to dependents. Failure policy decides whether to stop, skip or request a new plan version.
Technical explanation
The moving parts.
Plans are rows, not a running process. Each plan version is stored with its steps. State lives in PostgreSQL, so any worker can continue the work.
Transitions are atomic. A step's result and the steps it unlocks are committed together, so a crash cannot leave a half-finished join.
Delegation derives authority. When a step is scheduled, a child Mandate is created from its parent and checked to be a subset.
Agents share through records. A step reads the recorded outputs of its dependencies, and each handoff can be inspected.
Related capabilities
Have a multi-agent process that keeps breaking?
Tell us about it. We are looking for early access teams with real orchestration problems.