Introducing Kestrel
The operating system for production agents
Kestrel gives agent workflows durable execution, governed tool access, human review points, recovery, and evidence. It turns promising agent behavior into a system your team can inspect, improve, and run.
A good answer isn’t a production agent
AI models can draft, summarize, compare, and recommend. Production work needs approved system access, durable state, clear handoffs, human oversight, recovery paths, and proof of what happened. Kestrel provides the operating layer around agent work.
What Kestrel helps you do
Define workflows before launch
Map ownership, connected systems, allowed actions, review points, failure paths, and launch criteria before agents touch real operations.
Govern system access
Connect agents to real systems through typed tool contracts, allowlists, expected errors, and retained evidence.
Run with visibility and recovery
Durable execution with visible steps, saved state, checkpoints, approvals, retries, recovery, and final output history.
Evaluate real agent behavior
Use Ruhroh to run realistic tasks, preserve full work traces, review outcomes, and compare runs without relying on vibes.
Transfer operating capability
Deliver workflows, controls, runbooks, evaluation loops, and operational knowledge your team can keep using after launch.
The platform
One runtime for governed agent work
The foundation
Kestrel Runtime
The runtime behind each workflow: step execution, model and tool IO, durable state, checkpoints, approvals, recovery, persisted events, and replay material. Every capability below builds on the same execution model.
[01]
Kestrel Studio
Shape the workflow before launch: ownership, systems, allowed actions, review points, and launch criteria.
[02]
Tool Fabric & Connectors
Govern how agents reach enterprise systems through approved actions, expected errors, tool policies, and auditable access.
[03]
Ruhroh
Real tasks, full work traces, outcome reviews, and comparable runs so teams can see what agents actually did and what improved.
[04]
Operating Evidence
Capture context, tool calls, approvals, artifacts, outputs, checkpoints, and final state so each run can be inspected.
[05]
Developer Kit
Server-side SDKs, route helpers, CLI workflows, observability wrappers, and runner-service APIs for embedding Kestrel into existing architecture.
Operating evidence
Nothing disappears into a prompt
Every Kestrel run is visible while it executes and inspectable afterward. Actions, tool calls, approvals, artifacts, and final state become evidence your team can review, explain, and improve.
tool: crm.quotes allow: read, draft deny: send, delete inputs: account_id (required) errors: not_found, stale_record review: discount > 8% -> human approval evidence: request, response, actor, time
Kestrel ships inside Lumi delivery work. Standalone platform access is currently available through private preview.
Request demo accessOne runtime around your workflow. Five operating surfaces.
Before / after
Before Kestrel
- A promising prompt chain
- Scattered logs
- Unclear system access
- Informal approvals
- Limited validation
- A fragile handoff
After Kestrel
- A named workflow owner
- Governed system access
- Human review points
- Ruhroh evaluation loops
- Visible runs with retained evidence
- Operating guidance your team keeps
In the field
Quote review
Evaluate account history, pricing rules, and approval thresholds before recommendations become customer-facing.
Work-order triage
Prioritize work, identify missing information, coordinate assets, and retain evidence across operational decisions.
Healthcare intake & routing
Validate incoming requests, surface missing information, route safely, and separate recommendations from actions.
Dispatch exception handling
Coordinate disruptions, maintenance, compliance, and service commitments through approved workflows.
Engineering maintenance
Analyze defects, propose changes, execute validation, and retain a run history for every release.
Questions
Asked by every serious evaluator.
How do we get Kestrel?
Kestrel is included in Lumi's AI strategy and delivery work, from planning the first implementation through building and expanding it. Standalone platform access is currently available through private preview.
What stacks and models does Kestrel support?
Kestrel works with the model providers, cloud platforms, code repositories, and business systems your organization has already approved. You do not need to migrate to a new cloud or commit to one model provider.
How does Kestrel handle security and data?
Kestrel is designed to run inside your cloud and security boundary. Prompts, outputs, work history, and saved state remain in your environment under your identity, access, and data policies.
How is Kestrel different from open-source agent frameworks?
Agent frameworks help developers create agents. Kestrel helps teams run them in everyday business operations by managing long-running work, limiting access to approved systems, bringing people into important decisions, recording what happened, recovering from failures, and supporting repeatable testing.
How does Kestrel test whether an agent is improving?
Kestrel runs agents against realistic tasks, records how they completed the work, reviews the outcomes, and compares results over time. Teams can see what changed and whether the agent is becoming more reliable.
Can our engineers build on Kestrel without Lumi?
Yes. Kestrel provides server-side libraries, route helpers, command-line tools, monitoring integrations, and APIs that your engineers can use. Platform access is currently available through private preview.
What happens after Lumi hands off?
Your team keeps the working software, operating guides, run history, test suites, and engineering practices established during the engagement. Kestrel can continue running the AI agents your organization chooses to keep using.