GitHub ↗Book a pilot

§ structural verification for agentic AI systems

Maxwell.

You've deployed agents you cannot formally reason about. Maxwell changes that and proves the architecture containing them holds — before you ship.

Runtime monitoring sees what happened. Maxwell proves what cannot happen.

Get started on GitHub →Book a pilot
∎ deterministic analysis · zero runtime probes

§ The problem

The failure modes
are structural.

For decades, software was deterministic. Agentic systems replace predictable algorithms with probabilistic engines — wrapped in architectures designed for a different world. These aren't bugs you find in testing. They're architectural gaps that compound in production.

AG-002 · AG-004
Runaway Loop
Agent retries indefinitely. No exit condition. No timeout. Cloud bill uncapped.
AG-005
Blind Trust
LLM output fed directly to SQL, API, or financial decision. No validation in path.
AG-006 · AG-017
Black Box
Agent makes consequential decision with no log. No audit trail. Uninsurable.
AG-001
Naked Call
LLM call fails with no error handler. User sees a raw stack trace.
AG-012
Model Drift
Unpinned model version. Provider updates silently change behaviour.

§ The platform

Built on
Truss.

Maxwell uses the Truss platform — a rigid, layered architecture that treats your agent system the way a compiler treats code. Not a heuristic. Not a probability. A proof.

Invar extracts.
Axiom proves.
Verity certifies.
Provenance records.
Maxwellproduct personality
Truss Platform — rigid structural framework
Invar
Extraction engine
Extracts structural facts from your codebase. Dimensional stability — same input, same facts, always.
Axiom
Proof engine
Proves invariants hold across every execution path via CFG dominance analysis.
Verity
Certification engine
Produces the deterministic, machine-checkable result. Truth from proof.
Provenance→Verifiable audit artefact for auditors, regulators, and insurers.

§ The scan

Every violation is a failed
structural proof.

Maxwell scans your codebase in seconds. Not a warning. Not a suggestion. A deterministic finding with a named invariant and a regulatory mapping.

— named invariant
— file + line reference
— severity classification
— regulatory mapping included
△ /│\ /\│/\ / • \ maxwell v2.0.0 · invar/verity /───────\ deterministic verifier for agentic systems
targetsagecp
loadsource 30 files → SIR ok
buildCFG · dominance ok
engineInvar: IR annotation [s3.d7.r3.a2]
engineAxiom: invariant matching 17 rules
engineVerity: witness verification (52 signals)
VIOLATIONS (52) 6 × CRITICAL 8 × HIGH 28 × MEDIUM 10 × WARNING
backend/app/brokers/llm_client.py
L38[AG-004]Network call missing timeout. Infinite blocking detected.
L66[AG-001]Naked LLM call detected. Wrap in try/except block.
L66[AG-012]Model version via variable — not verifiable statically.
backend/app/routers/advisory.py
L281[AG-007]Empty exception handler. Silencing errors hides failures.
L235[AG-036]No explicit policy check in autonomous execution path.
OSS Cohort Percentile — vs Top 45 OSS Agentic Baseline
Guardrail Density
84thStrong
Decision Boundary
51thModerate
Replayability
2thLagging
Observability
40thModerate
Data Provenance
22thLagging
Policy Alignment
2thLagging
∎ verification complete · maxwell_audit.pdf

§ The registry

The taxonomy
that compounds.

49 invariants. 6 dimensions. Every agent architectural failure mode, named and formalised. The AG- registry is Maxwell's foundation — and it grows with every codebase audited.

17.
Active rules
Full verification set covering guardrails, observability, replayability, policy, and human oversight.
5+.
Regulatory frameworks
EU AI Act, DORA, NIST AI RMF, ISO 42001, US Executive Order 14110 — each invariant carries a full mapping.
49.
Invariant taxonomy
Tier S (decidable), Tier H (heuristic), and Tier R (runtime-hybrid) — every known agentic failure mode formalised.
∞.
Compounding corpus
Every scan sharpens detection. Competitors start from zero every time. Maxwell's corpus grows with every codebase audited.
Browse the full invariant library →
§ For developers

Start with Maxwell free.

Scan your agent codebase in under 5 minutes. No account required. Docker pull and run.

Get started on GitHub →
§ For regulated enterprises

Building in a regulated industry?

We run structured pilots with enterprise teams. Your codebase. Our engine.

Book a pilot →