Priority queues
Urgency, impact, confidence, risk, deadline and dependency determine order across both pipelines.
Explore SEOForge
Workflow Supervisor
Autonomous work stays prioritized, non-duplicative and inside budget.
The Supervisor coordinates both pipelines without becoming a production backdoor. It reserves capacity, watches agent heartbeats, resolves overlapping page scope, applies retry policy and enforces human approval.
Nothing publishes without approvalPaid product accessNo ranking guarantees
queued
12running
3approval
5
12running
3approval
5
CG-441 heartbeat · 14s ago
SO-731 retry 1/2 · provider 429
conflict · /guides/oauth locked
budget · 28,400 / 40,000 credits
Control plane
Every assignment has an owner, allowance and terminal state.
Budgets and usage
Capacity is reserved before execution; exhausted work stops safely instead of creating silent overages.
Heartbeats
Leases and progress events expose stalled workers without duplicating the assignment.
Conflict resolution
Overlapping URL, topic or repository scope is serialized, merged or returned for a decision.
Retry handling
Only transient failures retry, with bounded attempts, backoff and preserved evidence.
Approval enforcement
No agent or retry can cross the authenticated human production boundary.
Scheduling logic
The highest-value safe work moves first.
- 01
Receive scheduled, provider-triggered or human-requested work
Evidence, owner and next state remain visible.
- 02
Check duplicate fingerprints, page locks and dependencies
Evidence, owner and next state remain visible.
- 03
Score impact, confidence, urgency, effort and production risk
Evidence, owner and next state remain visible.
- 04
Reserve site, provider and model capacity from the paid allowance
Evidence, owner and next state remain visible.
- 05
Assign one specialist with a lease and heartbeat deadline
Evidence, owner and next state remain visible.
- 06
Retry transient failure or route terminal failure with evidence
Evidence, owner and next state remain visible.
- 07
Release capacity and require approval for any production proposal
Evidence, owner and next state remain visible.
Conflict example
Two agents want the same guide. One reviewable change survives.
- Content Growth requests a new OAuth comparison section
- Search Optimization requests a refresh after confirmed decay
- Supervisor detects identical preferred URL and overlapping intent
- Search finding becomes the parent assignment; new research is attached
- One combined proposal enters approval with one capacity ledger
Failure example · provider rate limit
Attempt 1 records the 429 and releases the worker. Backoff schedules one retry inside the reserved budget. A second transient failure closes the run as blocked and preserves partial evidence; it does not loop, overspend or publish incomplete work.
Usage controls by plan
Plans cap sites, tracked queries, provider operations, model work, content, media and concurrency. The ledger shows reserved, consumed and released capacity.
Human authority
Schedules can start research and preparation. Only an authenticated reviewer can approve the exact GitHub or WordPress production proposal.
Questions
Know the boundary before connecting production.
Can two agents create duplicate articles?+
Duplicate fingerprints, topic ownership, active assignments and URL locks are checked before work begins.
Can agents overspend?+
Runs reserve capacity first and fail closed when an allowance is unavailable.
What if an agent stops responding?+
Its lease expires, partial evidence is preserved and bounded recovery policy decides whether to retry or fail.
Can a retry publish work?+
No. Retry policy cannot change the production approval boundary.
How is priority decided?+
Impact, confidence, urgency, effort, risk, dependencies and paid capacity remain visible in the score.
Paid deployment · human authority
Put an accountable supervisor above both autonomous pipelines.
Book a paid product walkthrough to inspect queues, budgets, conflicts, retries and approval enforcement.