Coordinate specialized agents, deterministic steps, reviews, retries, and human decisions without losing the project context or the execution trail.
Input
Start from a card, document, trigger, or explicit mission.
Specialize
Assign focused steps to the right agent or deterministic tool.
Coordinate
Sequence dependencies or run independent work in parallel.
Review
Add correction loops, quality checks, and readable failures.
Verified output
Result, sources, and state return to the card.
Human approval
Protected actions wait for the right decision.
A pipeline defines who acts, what each step receives, when work can continue, and where a person must decide.
01
Start from a card, document, trigger, or explicit mission.
02
Assign focused steps to the right agent or deterministic tool.
03
Sequence dependencies or run independent work in parallel.
04
Add correction loops, quality checks, and readable failures.
05
Pause consequential changes until the right person decides.
Mission STL-128
Start from a card, document, trigger, or explicit mission.
Checklist
Linked context
Separate planning, production, review, and decision responsibilities instead of using one oversized prompt.
A production workflow needs project context, identities, tool boundaries, stop conditions, retries, approvals, and a result that returns to the source of truth.
Separate planning, production, review, and decision responsibilities instead of using one oversized prompt.
Follow the current step, prior outputs, failures, review decisions, and final state from the project.
Insert explicit approval before sensitive actions, external publication, or irreversible changes.
| Action | Human | Supervised | Approval |
|---|---|---|---|
| Read and analyze | Optional | Automatic | Automatic |
| Generate a draft | Review | Automatic | Proposal |
| Change project state | Decision | By policy | Proposal |
| External or sensitive action | Required | Protected | Required |
Agents are useful when a step requires interpretation. Pipelines remain useful for ordering, dependencies, routing, retries, and stop conditions.
Combining the two prevents every workflow decision from being delegated to a model. The agent handles the ambiguous part while the pipeline preserves the operating contract.
An orchestration is incomplete if its output stays in a terminal, an automation log, or a private chat. The result, failure, or approval must update the card and remain readable to the team.
That closed loop makes the next mission start from current project state instead of reconstructed context.
Manual coordination
Stellary orchestration
Handoffs
Manual coordination
Copy output between chats and tools.
Stellary orchestration
Pass structured outputs through pipeline steps.
Dependencies
Manual coordination
Remember what must finish first.
Stellary orchestration
Encode sequence, branches, and validation.
Failures
Manual coordination
Discover missing output late.
Stellary orchestration
Keep failed runs and stop conditions visible.
Approval
Manual coordination
Ask for review in another channel.
Stellary orchestration
Pause the workflow at an explicit human gate.
AI agent orchestration coordinates agents, tools, context, dependencies, and decisions so a multi-step objective can be executed as one controlled workflow.
No. Use deterministic steps for predictable transformations, routing, and checks. Use agents where interpretation, generation, or adaptive tool use is valuable.
Yes. Agents can have different roles, tools, rules, models, and project assignments. Pipelines can sequence or coordinate their contributions around one card or project outcome.
Require it before irreversible, external, financial, security-sensitive, compliance-related, or broadly scoped changes. Routine reversible work can use supervised execution after validation.
Start with a short sequence, one review step, and one explicit approval gate.
Create a Stellary project