Orchestrator router
Routes goals to specialist agents, merges their findings, and keeps each run aligned to the requested outcome.
AgentOps Studio · portable multi-agent ops lab
Run agent workforces from goal to approved outcome.
Coordinate specialist agents, cited RAG, an MCP sandbox, approval gates, and run traces. The public demo is native Node and Python behind Caddy. Compose and Kubernetes notes stay optional. This is AgentOps Studio, not the Contabo Agent Fleet product.
OrchestrationFastAPITF-IDF RAGMCP sandboxRun tracesApproval gates
4Demo workflows
6Specialist agents
6MCP tools
4Knowledge docs
Run AO-204Executive daily brief
PreviewOrchestrator chat
Agent execution graph
Goal
Research
RAG
Brief
Approval
Run inspector
TraceLive spans
ToolsMCP sandbox
ModeDeterministic
ApprovalSlack draft
Backlog1
Running2
Approval3
Done4
Runtime architecture
What this demo actually runs.
Orchestration, sandbox tools, cited retrieval, approval gates, traces, and a native public-host path. Optional Compose and cluster notes are labeled as scaffolding.
DAG workflow engine
Breaks complex work into ordered specialist steps with shared context, citations, and approval gates.
Kanban execution board
Turns every run into visible tasks with backlog, running, approval, failed, and done states.
Research and citations
Combines sandbox web search, document retrieval, and source-linked answers.
RAG memory layer
Indexes policies, playbooks, and runbooks into retrievable chunks with citations.
MCP tool registry
Connects tool handlers through permissioned templates, schemas, and sandbox invoke.
Trace waterfall
Shows agent steps, tool calls, RAG hits, artifacts, and approvals in one timeline.
Evaluation scorecard
Product research attaches a deterministic scorecard for completeness, citations, and sandbox safety. Every run still stores traces. There is no separate benchmark suite.
Public demo hosting
Native Node and Python on ports 3010 and 8010, behind Caddy. Compose and Kubernetes docs stay optional.
Specialized agents
A workforce model built around real operating roles.
Each agent has a defined responsibility, permitted tools, memory sources, output schema, and traceable handoff.
Orchestrator
Plans the run, delegates work, resolves dependencies, and synthesizes final outputs.
Deep Research
Calls sandbox web search. SearXNG is optional; otherwise the tool returns a deterministic demo fallback.
Knowledge Analyst
Retrieves policy, product, finance, and operations context from the RAG memory layer.
Tool Operator
Prepares sandbox Gmail and Slack drafts. The public demo never sends real mail, Slack, or GitHub writes.
Compliance Reviewer
Checks workflow outputs against policy, risk rules, and approval requirements.
Workflow Evaluator
Attaches a deterministic scorecard for completeness, citations, sandbox safety, and offline cost.
Workflow execution
Goal in. Agent graph out. Every step observable.
A user goal becomes an ordered specialist run. Agents execute one after another, pause for human approval when the workflow requires it, and produce evidence-backed artifacts.
Goal intake
The orchestrator turns a business goal into a structured run plan with constraints, expected artifacts, and success checks.
Agent graph
Specialist agents run in a fixed order and share one run context: plan, retrieval, research, compliance, and drafting.
Tool and memory
Agents call MCP tools, query RAG memory, collect citations, and attach every tool result to the run trace.
Approval and artifact
Risky actions pause for human approval, then the system produces a cited brief or sandbox draft.
Business workflows
Four workflows you can run in this demo.
Each one is a live specialist DAG. Start it from Workflows or the Dashboard. Other business workflows are not wired up here.
01
Executive daily brief
Cite support and ops docs, draft a Slack brief, and wait for approval. Seeded on first visit.
02
Support triage
Retrieve the refund policy and draft a careful reply. Approval required before the sandbox action.
03
Product research
Combine RAG playbooks with sandbox web research into a cited memo. Completes without an approval gate.
04
Compliance review
Check a proposed action against policy docs and return a risk note with citations.
Research and RAG
Ground answers in the seeded knowledge base.
Retrieval is TF-IDF over the markdown in demo-data/knowledge. Citations are attached to workflow steps. SearXNG is optional; without it, web search returns a deterministic fallback. Firecrawl and pgvector are not running.
Seeded markdownTF-IDF chunksCitation excerptsSandbox web fallbackOptional SearXNG
01Goal plannedOrchestrator picks the specialist order
02Knowledge retrievedCited chunks from the seeded markdown
03Sandbox toolsWeb search fallback and draft actions
04Artifact readyBrief or memo, approval when the workflow requires it
Runtime builder and MCP
Workflows, tools, knowledge, and approvals are already wired.
A visual builder is not part of this demo. Change behavior by running the live workflows, querying knowledge, and invoking sandbox MCP tools.
Runnable workflows
Four specialist DAGs are live: executive brief, support triage, product research, and compliance review.
MCP sandbox registry
Invoke knowledge search, web search, and sandbox Slack, Gmail, and GitHub handlers. Nothing is sent externally.
Seeded knowledge
Markdown policies and runbooks are indexed at startup. Queries return cited chunks.
Approval gates
External-style actions pause until you approve. Approval records a sandbox Slack post only.
Built with
The stack behind the public demo.
Next.js and FastAPI run the demo natively. Vector databases, Redis, and Kubernetes are documented options, not services this host depends on.
Next.js 16React 19FastAPIDeterministic DAGTF-IDF RAGMCP sandboxInternal tracesCaddyDocker Compose (optional)Kubernetes docsTerraform blueprints
Contabo public demo
Native Node + Python. Web :3010, API :8010, Caddy TLS on agentops.169.58.185.43.sslip.io. Distinct from Agent Fleet.
Native local dev
Hot reload with scripts/dev-api.sh (:8000) and npm run dev:web (:3000). No Docker required.
Compose (optional)
Same host ports 3010/8010 when Docker is available. Postgres, Redis, and SearXNG are optional profiles, not the live demo.
Cluster docs only
k3d manifests and Terraform READMEs are scaffolding. They do not provision the public demo.
Observability and evaluation
Debug a run from its own spans.
Every orchestrator, agent, tool, RAG, and approval step is stored on the run. The demo does not export those spans to Langfuse, and it does not run a separate benchmark suite.
Open the console and approve a seeded run.
Dashboard, Workflows, Runs, Knowledge, MCP, and Traces talk to the API. The first visit already has a cited executive brief waiting for sandbox approval.