Fisemarco / INTELLIGENCE CONTROL PLANE

Every agent step uses only the intelligence it actually needs.

Fisemarco separates agent execution from model execution. Use rules, memory, tools, local AI or frontier models when each is sufficient to complete the step reliably.

Explore Intelligence Control ↗See how it works ↓

Identity establishes who. Authority establishes what. Intelligence Control chooses how.

YOUR EXISTING AGENT↓Fisemarco INTELLIGENCE CONTROL

What does this step actually need?

Check invoice status → tool

↓

00Rules01Context / memory02Tools03Capabilities04Local AI05Private AI06Cloud AI07Frontier AI08Human

Illustrative routing; no providers are called.

Policy → safety → quality → privacy → reliability → latency → cost

STEP-LEVEL EXECUTION

One agent. Different intelligence.

Explore the routing of an enterprise sales workflow. These recommendations use the same decision engine as the runtime API, with declared example tools and model capabilities. No providers are called and no production results are implied.

EXAMPLE / STEP 01

Find target account

Task type: lookup
Route: tool
External model: Not required
Why this route: TOOL_MATCH
Quality requirement: 95% declared threshold

7example steps1external routes14.3%example dependency

Derived from example resource declarations. Routing recommendations, not model invocations or measured savings.

THE PROBLEM

Agents are becoming model-call machines.

Many workflows call a model for each step, including database lookups and known calculations. Each unnecessary call adds latency, external dependency and data exposure.

EVERY STEP USES A MODEL

Lookup → model
Relationship → model
Scoring → model
Objection → model

MINIMUM SUFFICIENT INTELLIGENCE

Lookup → tool
Relationship → memory
Scoring → capability
Objection → frontier

HOW IT WORKS

Intelligence becomes an execution decision.

Request an outcome

Your existing agent submits an individual step, its requirements and available resources.

Resolve what is known

Check validated context, memory, tools and versioned capabilities.

Apply the constraints

Enforce quality, action permissions, privacy, residency and human oversight.

Execute the chosen route

Your runtime executes a permitted callback or approved model with a bounded work permit.

Measure and improve

Report the outcome. Evaluate repeated patterns before promoting a reusable capability.

CONTROL BEFORE OPTIMISATION

Start by observing. Enforce when ready.

Observe records the actual route. Shadow adds a recommendation while preserving execution. Enforce selects the route. Privacy and human-approval constraints apply in every mode.

Run in Shadow Mode ↗

Personal or restricted data Local by default

Strategic planning Approved frontier model

Financial transfer Exact-task human approval

No sufficient approved route Explicit escalation

VALIDATED REUSE

Repeated reasoning can become a capability.

Repeated successful model tasks create review candidates. A developer implements a versioned handler, tests it against a dataset and safety cases, and an administrator reviews the evidence before promotion. New versions are candidates again. Withdraw a version to stop reuse.

CONTEXT MINIMISATION

Relevant facts. Smaller exposure.

Relevant context stays in your runtime. External providers receive selected fields with sensitive entries excluded and common identifiers redacted. Fisemarco stores routing metadata, hashes and reported outcomes; it does not store raw context, outputs or provider secrets.

Your context engineering and memory systems remain the sources of knowledge. Fisemarco selects the execution resource.

MEASURE THE REAL WORK

Intelligence Mix. Cloud Dependency. Quality.

Completed step executionsFrom your tracesNo-LLM executionFrom your tracesLocal intelligenceFrom your tracesCloud dependencyFrom your tracesFrontier dependencyFrom your tracesContext kept localFrom your tracesSuccess by routeFrom your tracesReported costFrom your traces

Organisation dashboards aggregate real traces on the backend. Unknown costs remain unknown. Savings are labelled estimates against your configured baseline. Shadow recommendations are not claimed as realised savings or equivalent quality.

Open your intelligence dashboard →

WORKS ALONGSIDE YOUR HARNESS

Keep your agent. Add control.

The Node SDK and signed REST API work with custom agents and orchestration callbacks. A local, text-only OpenAI-compatible bridge supports non-streaming chat completions with an explicit task mapping and validator. Ollama, vLLM and OpenAI-compatible provider adapters keep credentials in your runtime.

LangGraph, LangChain, CrewAI, Microsoft and MCP orchestration can call this SDK or REST boundary. Framework-specific integrations are not independently certified.

Integration guide → · SDK bundle

import { IntelligenceControl } from './intelligence.mjs';

const control = new IntelligenceControl({
  client: ftai,
  handlers: { 'crm.lookup': lookupAccount },
  models: { 'local-writer': localWriter },
  outboxPath: './private/intelligence-outbox'
});

const result = await control.run({
  step, context, grant, request
}, { validate: validateOutput });

External intelligence is optional until required.

Rules, memory, capabilities and local models can run without a cloud model. Tools still need their own connectivity. Routing currently requires the Fisemarco service; unavailable routing fails closed. A missing or failed provider yields an explicit error and trace rather than an unapproved fallback.

Connect the agent you already have.

Configure your organisation, register its runtime, issue bounded authority, and start in shadow mode.

Connect an agent ↗Configure Intelligence Control →