Catalog/Classes/Maxwell
CLASS · 03 / 06v0.4Live · v0.4

Maxwell

From a long view of the history of mankind, seen from ten thousand years from now, there can be little doubt that the most significant event of the nineteenth century will be judged as Maxwell's discovery of the laws of electrodynamics.

Named afterJames Clerk Maxwell (1831 — 1879) — physicist who unified electricity, magnetism, and light into four equations, and invented a demon to prove that information is a form of power.

99.3% pipeline completion2,800 cross-departmental runszero lost to handoff failure
01 — The science

Why we named it Maxwell.

Before Maxwell, electricity was one science, magnetism was another, and light was a third. They each had their own vocabulary, their own practitioners, their own laws. They were not obviously related. Maxwell looked at all three and wrote four equations. In those four equations, electricity and magnetism and light collapsed into a single field. Not three phenomena making an uneasy truce — one phenomenon, seen from different angles.

That is orchestration at its deepest level. Not coordination between things that remain separate. Unification into a system that behaves as one.

But Maxwell gave ALOFT something else, too. In 1867 he proposed a thought experiment: imagine a tiny intelligent gatekeeper sitting at the partition between two chambers of gas. It watches every molecule. When a fast one approaches from the left, it opens the door. When a slow one approaches, it stays shut. Without doing any apparent work, it sorts the system — concentrating energy, reducing disorder, apparently defeating entropy. This being became known as Maxwell's Demon.

The resolution came a century later: the demon does work. Not mechanical work — informational work. The act of observing, deciding, and routing has a thermodynamic cost. Erasing the demon's memory to make room for the next decision is where the entropy bill is paid. The demon is not magic. It is intelligence applied to routing — and intelligence costs something.

The Maxwell agent is that demon made operational. It sits at the junction of your pipeline. It watches what each agent returns, reads the state of the mesh, and routes with precision — not randomly, not round-robin, but because it knows which agent is best positioned for the next step. The difference between a coordinated enterprise AI system and a chaotic one is exactly this: the gatekeeper who turns information into decisions before the entropy compounds.

  • 01"On Governors" (1868) → control loops with stability conditions: each pipeline runs a feedback governor on health signals, with explicit damping to prevent retry storms.
  • 02Field unification → one handoff-contract schema: as Maxwell unified electricity, magnetism, and light into one system, one schema spans all departments.
  • 03Maxwell's demon / Landauer → observation has a cost: monitoring and memory erasure carry real thermodynamic cost; sampling budgets are first-class.

Read the design research →

Maxwell class

One scientific epoch ended and another began with James Clerk Maxwell.

— ALBERT EINSTEIN
Maxwell class
1 / 4Why is it called Maxwell?

One scientific epoch ended and another began with James Clerk Maxwell. The demon sorts the system — reducing disorder before entropy compounds.

James Clerk Maxwell
tap to continue →
Field Evidence
ACADEMIC BASIS
"Why Do Multi-Agent LLM Systems Fail?" (MAST)1,600+ annotated traces across 7 frameworks, 14 failure modes in 3 categories, κ = 0.88. Key finding: base-model improvements alone "will be insufficient to address the full taxonomy" — failures are coordination-structural, not model-capability.
Cemri et al. · NeurIPS 2025 D&B Spotlight
IN PRODUCTION
AWS us-east-1, Oct 2025 — recovery by governorsDNS race condition → cascading DynamoDB failure → synchronized retry storms (~11M failed requests) → congestive collapse. Recovery via manual throttling; remediations = velocity controls + queue-size-based rate limiting.
AWS postmortem · The Register, Oct 2025; InfoQ, Nov 2025
BENCHMARK
79% of multi-agent failures are coordination failuresSpecification (~42%) + inter-agent misalignment (~37%) dominate across popular frameworks; task verification accounts for the remaining ~21%.
MAST corpus · arXiv:2503.13657
02 — Agents in this class

Prototype agents.

Every class ships with reference agents calibrated to operational use cases. Fork them, deploy them, or use them as a template.

Procure-to-Pay Pipeline

cycle time −41% · 73% touchless

Hire-to-Onboard Coordinator

98.4% day-one ready

Order-to-Cash Orchestrator

DSO −6.2 days

Audit Readiness Orchestrator

evidence effort −78%

Quote-to-Contract Pipeline

9.4d → 2.1d
12-WEEK BETA · 9 DESIGN PARTNERS · 47,000 SHADOWED RUNS
03 — Qualification gate

The ALOFT pipeline, applied.

Every agent in this class passes the same five-stage gate. Below: the criteria specific to Maxwell agents at each stage.

ALOFT
01 · Curation
02 · Staging
03 · Deploy
04 · Operate
05 · Generalise
A→L→O→F→T
01
Curation
Pipeline scope; sub-agent classes mapped
  • Pipeline scope defined
  • Sub-agent classes mapped
  • Rollback boundary explicit
02
Staging
DAG eval ≥95%; contracts validated
  • DAG eval ≥ 95% pass rate
  • Sub-agent contracts validated
  • Failure-mode replay
03
Deployment
Lineage published; scope ≤ tier-3
  • Lineage published
  • Rollback contract signed
  • Scope grant ≤ tier-3
04
Operation
Pipeline success ≥99%; drift = 0
  • Pipeline success ≥ 99%
  • Sub-agent drift = 0
  • Cost variance < 8%
05
Generalisation
Pattern to registry; new domain qualified
  • Pattern published to registry
  • Adapted to new domain
  • Playbook memo released

Ready to deploy a Maxwell agent?

Open the Workbench →Back to catalog
All classes
FeynmanFermiMaxwell· currentRamaMarcusWheeler