Reason
Agentic Teams
Orchestrate specialist AI agents into collaborative teams. Choose your topology, set budgets, schedule autonomous workflows. With QuoxMind Pro, a learning router favours the best-performing specialists. Powered by QuoxEngine.
In plain words
What it is, where it lives, when to reach for it
- What is it
- A dashboard feature of QuoxCORE that runs several specialist agents as one team under an orchestrator.
- Where do I use it
- In the dashboard, on the Agentic Teams page in the AI Studio section.
- When would I use it
- When one agent is not enough and a task needs several specialists working together under your rules.
- How do I use it
- Open Agentic Teams, click Create Agentic Team, then choose an orchestrator and at least one member agent.
QuoxCORE is the free, self-hosted platform underneath this. What is QuoxCORE
Human-in-the-loop
Real autonomy.
A hand on the wheel.
Agent teams plan, split the work and execute on their own, until a step crosses a threshold you set. Then the team holds and the decision routes to a person through the approval inbox.
Gate · policyHuman-in-the-loop gates
Require approval before a run starts or after each task, and escalate high-risk or unrecognised tool calls to a person. The check runs in the task service, not in a prompt.
▸ enforced in-engine · not advisoryDecide · HITL inboxApproval inbox
Pending decisions arrive with full context: what the agent wants to do, why, and the evidence behind the request. Approve, deny, or ask the agent to explain itself. No grepping logs to find out what is stuck.
▸ full context · one tapContain · scoped toolsScoped tool access
Autonomous runs are limited to a scoped tool allowlist and budget; anything outside it escalates to a person.
▸ scoped tools · escalatesLive demonstration
Watch a team run an objective.
Six agents from the Quox core registry, running the daily security audit as a team. An autonomous schedule fires, the orchestrator decomposes the objective, specialists hand the work along the line, and the run holds at the gate when a step needs a person.
An illustration of a team run, built from real agents: CommanderQ, PRISM, WARDEN, SCHOLAR, DAEDALUS and VULCAN are shipped Quox agents, and the gate and approval steps follow the shipped HITL flow.
When the human approves, watch the hands behind this page: the whole team lights up, the same governance a whole organisation gets from the Quox enterprise platform.
How it works
From objective to auditable output.
One flow: you set the objective, the orchestrator splits the work, specialists execute, and everything lands on the record.
Choose an orchestrator and specialist agents. Set policies, budgets and autonomy rules.
Describe what you want done: a high-level goal, not a script.
The objective is broken into concrete tasks and each one is assigned to the right specialist.
Monitor progress in real time. Approve plans, review checkpoints, or let autonomous schedules handle it.
Auditable output, monitored from the dashboard. Each objective produces a VOLT bundle and AEE envelopes for the record.
Topologies
Pick the shape of the team.
Four coordination topologies, one engine. Choose the structure that fits the work, not the other way round.
Capabilities
Everything you need for multi-agent teams.
Define, deploy and monitor autonomous agent teams from a single dashboard.
Team manifests
Define teams as JSON manifests: orchestrators, specialists, policies and budget limits. Collaboration you can version-control.
Objective decomposition
Submit a high-level objective. The orchestrator decomposes it into concrete tasks and assigns each one to the right specialist.
Autonomous scheduling
Cron-based scheduling with budget guardrails. Teams execute objectives on autopilot within configurable daily token limits.
Family-tree hierarchy
A visual tree of orchestrator-to-member relationships. See who does what at a glance.
Live objective tracking
Real-time SSE streaming shows objective progress, task status and agent activity as it happens.
Budget guardrails
Set daily token limits, tool call caps and a maximum for unattended objectives. A team’s schedule is disabled automatically after consecutive failed runs.
Team plugin packs
Bundle a team as a pack (JSON) and load it into your instance. Marketplace distribution is planned.
Shared context
Team instances share key-value context across tasks. Every agent works from the same working memory.
Topology selection
Choose how your team coordinates: star for flat dispatch, hierarchical for pod leads, pipeline for sequential stages, mesh for parallel work with a final synthesis step.
Learning router
With QuoxMind Pro, a learning router scores specialists and favours the best performers (90% exploit, 10% explore). Without it, tasks are spread round-robin.
Circuit breaker
After three consecutive failed scheduled runs (configurable), the schedule is auto-disabled so it stops spending tokens.
Define · share
A team is a file.
A team starts as a manifest: who leads, who specialises, when it pauses for a human, and what it may spend. Review it in a pull request, roll it back like code, or bundle it as a pack and load the same team into another instance.
Approval and budget policies sit right there in the file. They are not a suggestion in a prompt, they are configuration the task service enforces.
{
"id": "secops",
"orchestrator": { "agentId": "sentinel" },
"members": [{ "agentId": "cipher" }, ...],
"policies": {
"maxBudgetTokens": 200000,
"maxBudgetToolCalls": 80,
"requireApprovalBeforeStart": true,
"checkpointAfterEachTask": true
}
}Use cases
Teams built for real work.
Pre-configured team templates for common operational workflows.
01 · SecOpsSecurity audits
A SecOps team runs daily vulnerability scans across your fleet. This is the team from the demonstration above.
02 · DevOpsInfrastructure health
A DevOps team monitors, diagnoses and fixes infrastructure issues.
03 · DevelopmentCode review pipelines
A development team reviews PRs, runs tests and reports findings.
04 · AuditCompliance monitoring
An audit team tracks policy compliance and generates reports.
Where it stands
What's proven, what's not yet.
- Team manifests validate on load and the four topologies (star, hierarchical, pipeline, mesh) are real: TopologyBadge renders whatever manifest.topology.type declares
- Autonomous cron scheduling is live, with overlap prevention, timezone support and a configurable circuit breaker (default: 3 consecutive failed scheduled runs disables the schedule)
- Orphan-run recovery on service restart and a family-tree view of the agent hierarchy are shipped code, not concept art
- Budgets enforce daily token and tool-call caps, but per-request token-in accounting is incomplete (STATUS_AND_LIMITS_AUTOMATION.md)
- HITL routing, the inbox and SLA escalation are Beta throughout; the server-side VOLT-witness path for approvals has not been live-verified against the deployed containers yet
- The learning router (epsilon-greedy exploit/explore scoring) is real but gated behind QuoxMind Pro, and a fresh team’s Performance tab reads "No performance data yet" until objectives actually run: nothing has learned from live traffic yet
- Domain-compatible fallback routing exists as a helper but is not yet wired into failure handling
- Marketplace distribution of team packs is planned; packs load as JSON today
- The task-time skill loop (Crystal drafting from a run) is WARDEN-only; other team members do not draft Crystals from their own objectives yet
- No sub-agent creation flow: a team is assembled from the existing registry, not generated on demand
Autonomy you can hand to an auditor.
approvals routed to a personbudgets enforced in-engineVOLT bundle per objective
Many hands, one gate
Ready to build your first team?
Define objectives, assign specialists, set budgets, and let your agents collaborate with a human on the important calls.
Go deeper