The model is the commodity
Two teams pointing the same frontier model at the same task get wildly different results. The difference is the harness — tools, permissions, context, validation, and recovery.
Appearance
Skip to content
01 02 03 04 01 02 03 04 05 Healthcare IT · Medplum MCP
Applied AI · Enterprise
The team with the best harness wins — not the best model.
Agent harnesses, context engineering, retrieval, and fine-tuned open-weight models — built on infrastructure you own, with permissions, version history, and audit trails already in place.
The Thesis
Sell the work, not the software
Agents break the old SaaS bet. Customers stop buying a faster way to do the task and start buying the task done — on a substrate the agent can actually operate.
Autonomous is a liability claim
In regulated domains the durable design is human-in-the-loop. The defensible property is not "rarely wrong" — it is catchable and reversible.
Two API surfaces, not one
REST for developers, MCP for agents — over the same record, the same permissions, the same audit log. Humans and agents enter through different doors into one house.
Tokens are your cost of goods
When you sell the work instead of the seat, margin is whatever the harness does not burn re-reading context. Routing, caching, and early evals decide the unit economics.
Capabilities
What we actually build
The model is the easy part. The harness — tools, context, retrieval, evals, and serving — is where durable advantage accumulates.
01
Agent Harness Engineering
MCP tool layers, permission scoping at the boundary, output validation, error-recovery loops, approval and review surfaces, multi-agent orchestration, tracing and observability.
02
Context Engineering
Context budgeting, system and tool-prompt design, prompt caching, compaction and summarization, session plus long-term memory, self-describing metadata, and model routing — small models for routine steps, frontier only for hard calls.
03
RAG & Knowledge Systems
Ingestion and chunking, embeddings, hybrid BM25 + vector search, rerankers, graph RAG, grounded answers with citations, incremental freshness, and permission-aware retrieval so a query can never surface a document the caller could not open.
04
Model Training & Fine-Tuning
LoRA / QLoRA, supervised fine-tuning, DPO preference tuning, distillation of frontier behaviour into small cheap models, dataset curation and synthetic generation, quantization, and self-hosted serving on vLLM or Ollama.
05
Evals & Guardrails
Task-level eval harnesses, regression suites on real traces, faithfulness and recall scoring for retrieval, red-teaming, and cost and latency budgets enforced in CI.
06
AI Infrastructure & MLOps
GPU provisioning, on-prem and air-gapped deployments, model gateways with fallback, prompt and model versioning, and cost attribution per workflow.
Anatomy of a Harness
Five pieces around the model
Anyone can put an MCP server over an app in a weekend. Almost no one can hand an agent permissioning, version history, and a compliant review surface that already existed before the agent showed up.
Single system of record
Humans and agents act through the same record. No parallel backdoor, no split audit trail.
Self-describing metadata
Every entity is introspectable, so an agent can discover what exists and what it is allowed to touch.
Native permissions
Access controls enforced at the system boundary — the agent hits the same RBAC a human would.
Modelcommodity layer
Native version history
Every write is a versioned, reversible record by default. Error recovery is real, not aspirational.
Human review surface
A cockpit where an operator or auditor can see, approve, and reverse agent actions.
How We Engage
From audit to work sold as a service
012 weeks · fixed fee
AI Readiness Audit
Substrate assessment, agent-readiness score, use-case shortlist ranked by value over risk, and a token-cost plus unit-economics model.
026–12 weeks
Agent Build Pod
One production workflow shipped end to end: harness, MCP tools, evals, review surface, and a runbook your operators can own.
034–8 weeks
Model Lab
Dataset curation to LoRA to eval to self-hosted serving — with a measured quality and cost delta against the frontier baseline.
04Ongoing
Work-as-a-Service
We run the agents and are paid per unit of work completed, with SLAs and a human approval gate where regulation demands one.
Built On Our Own Substrate
The offerings that make the harness real
Inherit the substrate instead of spending a year re-implementing audit, RBAC, and version history.
01
Hydra AI / Hydra MCP
The operational brain over Frappe & ERPNext
▼
The world's most focused AI dev agent for ERPNext and Frappe — with deep ERP domain knowledge, MCP tool calling, and optimized token management. Hydra exposes patient operations, clinical orders, billing, inventory, accounting, and any custom object as a structured, permissioned, auditable tool surface.
02
Espresso Cloud
Where the agents and models actually run
▼
Private CI/CD, self-hosted GPU, monitoring, and disaster recovery. Spin up the environments your harness needs — including air-gapped and on-prem — with enterprise SLAs. The inference an agent burns is your cost of goods; the infrastructure has to be yours to optimize.
03
Fossible Sprints
The AI-augmented pod that builds the harness with you
▼
A dedicated dev pod — BAs, PMs, and engineers — delivering with an AI-augmented workflow and enterprise-grade guardrails. From spec writing and test-driven implementation to automated static analysis, AI code review, and auto-deployed PRs. Flat monthly fee, no surprises.
The clinical brain sits on the same substrate
Medplum MCP exposes the full FHIR R4 resource tree as semantically described tools, with SMART-on-FHIR scopes and AuditEvent generation enforced at the tool boundary. Pair it with Hydra MCP when the workflow spans operations and clinical data.
Explore Healthcare IT →Stack
Model-agnostic by design
Claude / GPTFrontier models
Llama · Qwen · MistralOpen weights
MCPAgent tool surface
LangGraphOrchestration
vLLMSelf-hosted serving
pgvector · QdrantRetrieval
PEFT / LoRAFine-tuning
LangfuseTracing
RagasEval suites
Who It's ForBuilt for teams that need
Built for teams that need
agents that survive production
Incumbent Operators
Years of data and distribution sitting behind an agent-hostile API — the asset is there; the substrate is not.
AI-Native Startups
Burning runway re-implementing audit logs, RBAC, and version history instead of shipping the work that is actually yours.
Healthcare & Payer Back-Office
Prior auth, claims, eligibility, denial management, and documentation — high volume, high audit, low tolerance for a black box.
Finance, Insurance & Lending
Rules-bound operational work that agents can carry, if every write is permissioned, versioned, and reviewable.
Regulated & Air-Gapped Workloads
You own the code, the data, and the deployment. Model-agnostic, open-weight capable, hostable anywhere.
Demo-to-Production Teams
The happy-path demo works. Production numbers do not — because the harness, evals, and unit economics were never built.
Get Started
Stop demoing the happy path.
Talk to us about the harness, the evals, and the unit economics — not another proof-of-concept that dies in production. We'll scope an engagement within 48 hours.