# Bitligence — intelligent agents you can trust to act

> Bitligence builds intelligent agents for consequential enterprise work and the runtime governance that makes them safe to act. The intelligence comes from the foundation model (a commodity); the runtime governance is the product Bitligence builds — every agent action is checked against its capability profile, active policy, context, and live system state, in the request path, before it executes. Synchronously in ~8 ms, fail-closed, with every decision signed and replayable.

Most "AI safety" controls are retrospective — model cards, dashboards, after-the-fact review. Bitligence sits in the path and can stop an unauthorized action before it happens. That is the difference between an agent that can answer and an agent you can trust to *act*. Contact: hello@bitligence.com.

## What Bitligence builds
- **Intelligent agents** for consequential work — the accessible, big-tent product.
- **Runtime governance** — the differentiator. The layer that makes a foundation-model-powered agent trustworthy enough to take real action. Trust is the product, not the model.
- Framed so "we build agents" never competes with OpenAI / Anthropic / agent frameworks — Bitligence governs their output, across every model and tool.

## Runtime governance (the core product)
Every action is evaluated in-band before it executes:
- **Capability profiles** — authorized *outcomes*, not raw tool lists.
- **Cross-agent session graph** — multi-agent context, so threats are caught at the aggregate, never just per call.
- **Action classes + effect annotations** — what a step actually does and how reversible it is.
- **Confidence-laundering & interpretation-divergence detection** — catches an agent overstating certainty or quietly drifting from intent.
- **Reversibility-gated human-in-the-loop** — high-stakes / irreversible actions stop and route to a human with the reasoning attached.
- **Policy-as-code + signed attestations** — versioned, diffed, rolled back; per-route and per-tenant.
- **Sentinel + Coordinator** — detection kept separate from recommendation.

### Key facts
- Enforcement: in-band (synchronous, in the request path), ~8 ms.
- Default on timeout: fail-closed (configurable per policy).
- Decisions: cryptographically signed and replayable; no sampling.
- Verdicts: permit / redact / deny — every one reconstructable for audit.

## Products
- **AO (Assurance Observatory)** — the runtime governance overview and oversight surface. The live demo on the homepage IS the product.
- **Keelson** — the build layer above agent frameworks. Describe an agent once as a single declarative spec; Keelson compiles it to run on LangGraph, CrewAI, or your own stack, with a deterministic backbone (predictable code by default, intelligence only where judgment is required). Change frameworks without rewriting the agent. (/keelson)
- **CX Agent** — the first horizontal agent. One governed customer agent — understand, resolve, escalate, attest — running the same job across every industry on the same runtime. (/cx)
- **Insurance — Atlas + Praxis** — two governed P&C claims agents. Atlas handles FNOL voice intake and captures a structured claim file during the call (≥95% field completion at first contact). Praxis runs the claim file after intake. Every decision is scored, traced, and routed to the adjuster's oversight inbox; high-stakes work never auto-executes. Integrates with Guidewire ClaimCenter and Duck Creek. (/insurance)

## Industries
- **Insurance** (live) — P&C claims (Atlas + Praxis), every decision scored to an oversight inbox.
- **Pharma & life sciences** — pharmacovigilance & safety operations (MedDRA, E2B R3, Veeva Vault).
- **Healthcare payers** — prior authorization & coverage review (HL7, FHIR, X12 278).
- **Fintech & lending** — origination, KYC/AML, fraud decisioning (Plaid, Equifax, OFAC).
- **Legal & compliance** — contract & obligation review (iManage, NetDocuments).

## One control point, many problems
The same in-path governor that arbitrates coordination also delivers: signed, replayable audit; prompt-injection and multi-turn crescendo defense (caught at the aggregate across the session graph); failure attribution (which agent, which step); collusion detection; and per-agent cost & latency budgets.

## Deployment & integration
- Deploy: cloud, private cloud, VPC, on-prem, or fully air-gapped — data never leaves your perimeter.
- Governs the output of any foundation model and any agent framework.

## Security & compliance
- SOC 2 Type II, ISO 27001, HIPAA, FedRAMP-ready, EU AI Act.
- Every permit / redact / deny is cryptographically signed and reconstructable.

## Links
- Homepage (with the live AO demo): https://bitligence.com/
- Keelson — the build layer: https://bitligence.com/keelson
- CX Agent — horizontal agent: https://bitligence.com/cx
- Insurance — Atlas + Praxis: https://bitligence.com/insurance
- Book a demo: https://calendly.com/bitligence/30min
- Contact: hello@bitligence.com
- Full site content for LLMs: https://bitligence.com/llms-full.txt