Lancelot is composed of modular subsystems gated by dependency-resolved feature flags. Governance is the foundation. Capability is progressively enabled on top of it. Every subsystem is independently kill-switchable.
20
Subsystems
5,500+
Tests Passing
5
Patents Filed
4
Risk Tiers
Design Principles
Governance
Governance over convenience. Every action is constrained.
Verification
Verification over speed. Outcomes are confirmed, not assumed.
Determinism
Deterministic context over retrieval. No hallucinated state.
Receipts
Receipts over trust assertions. If there is no receipt, it didn't happen.
Reversibility
Reversibility over irreversible autonomy. Every action has a rollback path.
Governance Foundation
The Constitutional Layer
The core primitives that make every other subsystem governed. Capability is built on top of these, not the reverse.
Constitutional Soul
Versioned constitutional document defining hard behavioral boundaries. Immutable without owner approval. Linted, validated, and immune to prompt injection. Enforced at pre-execution, runtime, and post-execution stages.
Progressive trust graduation with instant revocation. 50 consecutive successes triggers a graduation proposal. A single failure triggers instant revocation. Trust is earned slowly and lost immediately. Binary trust models are a liability at scale.
Immutable audit trail for every action, check, and outcome. Structured receipts record the governance chain: action, risk tier, Soul check, verification result, rollback reference. Both success and failure paths produce receipts.
Four-tier memory architecture: Core, Working, Episodic, Archival. All edits are atomic and auditable. The agent does not write to its own memory directly. Memory poisoning is architecturally blocked.
The mechanisms through which governed actions are planned, executed, and tracked. Every capability routes through governance, not around it.
Tool Fabric
Provider-agnostic execution with Docker sandbox and policy gating. Six mandatory security gates before any skill reaches production: manifest, static analysis, sandbox, enforcement, approval, monitoring.
Framework-level control over desktop and server apps across 9+ UI frameworks. 61 action types. 17 native MCP tools. Spatial Map Engine for 2D spatial understanding. 5-tier cascade control with Vision fallback. No APIs, no plugins, no screen reading.
Ephemeral sub-agent orchestration with monotonic governance restriction. Sub-agents can only be more restricted than the parent, never less. LLM-powered task decomposition into parallel execution groups. Five-step lifecycle: Decompose, Spawn, Execute, Intervene, Collapse.
Honest completion semantics across arbitrary time horizons. No simulated progress, no hallucinated completions. Plans track verified outcomes only. Database-backed durable execution that persists across hours, days, or weeks.
Governance-aware cron and interval jobs with approval gates. No job executes outside the governance pipeline. Supports deferred execution, recurring tasks, and dependency chains.
Six mandatory gates before any third-party capability is installed: manifest declaration, static analysis for banned patterns, Docker sandbox execution, runtime policy enforcement, owner approval, and ongoing monitoring with revocation.
How Lancelot connects to external services, other agents, and other Lancelot instances. Always through a governance proxy, never directly.
Governed Connectors
Proxy-mediated external integration with credential vault isolation. Connectors produce request specifications that pass through policy evaluation, risk classification, vault credential injection, execution, and verification. The agent never sees raw credentials.
Multi-instance governed topology with trust-verified peering, federated receipt chains, and centralized mission control. Instances share governance state without sharing raw data. Graph-based topology builder.
Google A2A standard implementation with governed inbound/outbound pipelines, agent registry with trust tiers, and agent card generation. All inter-agent communication passes through governance.
System health monitoring with liveness, readiness, and diagnostic state transitions. All transitions produce verifiable receipts. Cascading health checks across dependent subsystems.
Real-time visibility, forensic debugging, and one-click compliance reporting. All derived from the same immutable receipt trail.
Compliance Export
One-click SOC 2 Type II, ISO 27001, and GDPR Article 30 report generation from the receipt DAG with chain integrity verification. No manual evidence gathering. The receipt trail is the compliance evidence.
Pause, inspect, modify, and resume at any receipt in the execution DAG. Three modes: inspect (read-only), replay (re-execute with same inputs), fork (branch from any point with modified parameters).
Playbook engine with severity classification, automated containment actions, and post-incident receipt trails. Integrates with the receipt DAG for root cause analysis and forensic reconstruction.
