Harnesses
Each agent runs the harness its task needs, with shared orchestration, storage and review.
Self-hosted agent runtime
Stablehand runs specialist agents inside your product and workflows, with your tools, approvals and budgets, on infrastructure you control.
Specialist agents inside a SaaS product, on Stablehand
For a software company, we built the first prototype of an agent product and connected it to Stablehand, our self-hosted agent runtime. Specialist agents use company context and tools to produce work that people review.
The agent catalog covers customer research, planning and content production. Stablehand provides the orchestration layer; agent definitions and tools carry the business logic.
Claude Agents, LangGraph and PI on one runtime
Deployment architecture
Your product
Chat, an embedded workflow, or a background trigger.
Your systems
Business tools, files and approved data access.
Self-hosted orchestration
Claude Agents · LangGraph · PI Agent
Session workspaces
Run lifecycle and event replay
Permissions and human review
Budget policies
Approved external endpoint
Provider API through an explicit, controlled route.
Work to review
Artifacts, citations, run history and corrections.
Provider API calls leave through the gateway, under rules you set on what may leave.
Context and tools
Specialist agents get approved tools, domain instructions and scoped business context, and every output links back to its source material.
Run history and review
Every run, tool call and artifact is recorded. People review outputs and approve the steps that need judgment.
Budgets, access and routing
Persistence, model routing, access boundaries and inference budgets are set for each deployment.
Engagement scope
For a product team, that means a shared agent runtime. For an operating company, a workflow around intake, research or document review. For a group of companies, reusable infrastructure with different tools and permissions for each business.
Our enterprise work also includes data-platform architecture: ingestion, mappings, human checkpoints and integration with existing financial workflows.
First build
Choose the work
Name the user, the workflow and the decision the system must support.
Define the test
Use representative tasks and failure cases to establish a baseline.
Build and integrate
Connect the tools, run the agents and put review where the work needs it.
Measure and improve
Use observed failures to decide whether the next change belongs in data, tools, prompts or weights.
Work with us
Tell us what the work is, what data it touches and what constraints apply. We’ll work out what to build and how to measure it.