WorkingAgents + Accelsius: Agent Governance Meets the Physics of AI Infrastructure

By James Aspinwall, co-written by Alfred Pennyworth (my trusted AI) — March 7, 2026, 19:05

Accelsius cools the hardware that runs AI. WorkingAgents governs the AI that runs on the hardware. As data centers scale from megawatts to gigawatts to support agentic workloads, the cooling infrastructure itself becomes a mission-critical system that needs intelligent, governed management. This partnership sits at the intersection of physical infrastructure and AI governance — two layers of the same stack that are about to converge.

What Accelsius Brings

Accelsius is a liquid cooling company founded in 2022, commercializing two-phase direct-to-chip technology originally developed at Nokia Bell Labs. Series B closed January 2026 at $65M led by Johnson Controls with Legrand joining — strategic investors who are the largest building technology and electrical infrastructure companies in the world. Total funding: ~$89M+.

NeuCool Platform — Two-Phase Direct-to-Chip Cooling:

NeuCool MR250 — Row-Based CDU (Generally Available Oct 2025):

NeuGuard — Enterprise Support Program:

Intelligent Monitoring:

Customers and Partners:

Target Market: Hyperscalers, OEMs, colocation providers, and enterprise data centers deploying AI/HPC workloads at 50+ kW per rack densities.

What WorkingAgents Brings

WorkingAgents is an AI governance platform with three gateways (AI, Agent, MCP) that governs 60+ MCP tools with capability-based access control, audit trails, and guardrails on every agent action. The platform provides multi-step workflow orchestration, cost attribution, and observability across all agent interactions.

What WorkingAgents doesn’t have: physical infrastructure management, cooling telemetry, hardware monitoring, or data center operational technology. What it does have: the governance layer for autonomous systems that make decisions affecting mission-critical infrastructure.

The Strategic Thesis

Data center cooling is becoming an autonomous, AI-managed system. The industry trend is clear:

Accelsius already has intelligent monitoring with SNMP/IPMI/Redfish integration and auto-failover. The next step is AI agents that autonomously manage cooling infrastructure — adjusting coolant flow rates based on GPU utilization forecasts, preemptively routing cooling capacity before workload spikes, coordinating with power management systems, and making decisions that affect hardware worth millions per rack.

Those agents need governance. An autonomous cooling agent that makes a bad decision doesn’t just waste energy — it can thermal-throttle a training run that costs $100K/hour, or worse, damage GPU hardware that takes months to replace.

This is where WorkingAgents enters: governing the AI agents that manage the physical infrastructure that runs AI.

The Gap Analysis

Accelsius Gap WorkingAgents Solution
Monitoring is reactive — alerts and dashboards, not autonomous optimization Agent Gateway orchestrates multi-step cooling workflows with checkpoints
No permission model for who/what can adjust cooling parameters Capability-based access control scopes each agent and operator’s authority
No audit trail connecting cooling decisions to workload events Unified logging across infrastructure and workload layers
No guardrails preventing dangerous cooling configuration changes Pre-execution validation blocks unsafe parameter changes
No multi-system coordination (cooling + power + workload) Agent orchestration across infrastructure, power, and compute agents
No cost attribution linking cooling energy to specific workloads Token-level cost tracking extended to infrastructure cost per job
Single-system monitoring (cooling only) Cross-system observability across cooling, power, compute, and network
WorkingAgents Gap Accelsius Solution
No physical infrastructure management capability NeuCool platform with real-time cooling telemetry
No hardware-level monitoring or telemetry SNMP, IPMI, Redfish, DCIM integration
No data center operational technology expertise Decades of thermal engineering from Nokia Bell Labs heritage
No edge/remote deployment infrastructure Auto-failover for edge facilities with remote management
No hardware reliability or warranty infrastructure NeuGuard with $100K/rack coverage and CNA warranty backing
No relationship with hyperscalers/colo operators Equinix, Johnson Controls, Legrand partnerships

Synergy Areas

1. Governed Autonomous Cooling Management

The primary opportunity. AI agents that manage cooling infrastructure in real time, with WorkingAgents providing the governance layer:

Cooling Optimization Agent
  ✓ NeuCool: read thermal telemetry (all racks)
  ✓ NeuCool: adjust coolant flow rate (within safe bounds)
  ✓ NeuCool: adjust facility water temperature setpoint (±2°C)
  ✓ Workload Manager: read GPU utilization forecasts
  ✓ Alerts: send thermal warnings to NOC
  × NeuCool: disable cooling on active racks
  × NeuCool: change refrigerant type
  × Power: modify power distribution
  × Hardware: shut down compute nodes

Capacity Planning Agent
  ✓ NeuCool: read historical thermal data
  ✓ NeuCool: read CDU capacity utilization
  ✓ Workload Manager: read job queue and scheduling
  ✓ Reports: generate capacity forecasts
  × NeuCool: modify any operational parameters
  × Power: modify any operational parameters

The cooling agent can optimize within defined safety bounds — it can adjust flow rates and temperature setpoints but cannot disable cooling on active racks or make changes outside its authorized range. Every adjustment is logged with the thermal data that triggered it, the workload context, and the energy impact.

Human-in-the-loop for high-risk operations: “Agent recommends increasing facility water temp from 38°C to 42°C to save 15% cooling energy during off-peak — approve or deny?”

2. Cross-Layer Infrastructure Observability

Modern AI data centers have four interdependent layers: compute, network, power, and cooling. A problem in one layer cascades to the others. GPU utilization spike → thermal spike → cooling response → power draw increase → potential circuit overload.

WorkingAgents provides unified observability across all four layers:

{
  "event": "training_job_scale_up",
  "cluster": "gpu-rack-14-16",
  "timeline": [
    {
      "layer": "compute",
      "action": "GPU utilization increased 40% → 95%",
      "agent": "workload_scheduler",
      "timestamp": "2026-03-07T14:00:00Z"
    },
    {
      "layer": "cooling",
      "action": "NeuCool flow rate increased 60% → 85%",
      "agent": "cooling_optimizer",
      "guardrails": { "thermal_safety": "passed", "flow_rate_limit": "within_bounds" },
      "timestamp": "2026-03-07T14:00:03Z"
    },
    {
      "layer": "power",
      "action": "Rack power draw increased 45kW → 72kW",
      "agent": "power_monitor",
      "guardrails": { "circuit_capacity": "passed", "budget_check": "passed" },
      "timestamp": "2026-03-07T14:00:03Z"
    },
    {
      "layer": "cooling",
      "action": "CDU MR250 capacity at 88% — pre-staged adjacent CDU",
      "agent": "capacity_planner",
      "timestamp": "2026-03-07T14:00:05Z"
    }
  ],
  "total_cooling_cost_delta": "+$12.40/hr",
  "attributed_to": "customer: acme-corp, job: llm-finetune-v3"
}

One log traces the entire cascade from workload to cooling to power to cost. When something goes wrong — and in a 250kW-per-rack environment, the failure modes are severe — operators see the full chain of causation, not fragments scattered across four different monitoring systems.

3. Workload-Aware Cooling Cost Attribution

AI infrastructure operators need to know what cooling costs per job, per customer, per GPU-hour. Today, cooling is a facility-level overhead allocated by square footage or rack count. With NeuCool’s per-rack telemetry integrated through WorkingAgents’ cost attribution:

This transforms cooling from an undifferentiated facility expense into a metered, attributed, optimizable cost center — critical for colocation providers and cloud operators who need to price AI compute accurately.

4. Predictive Maintenance with Governed Response

NeuCool’s hot-swappable components (pumps, power supplies, control boards, sensors) combined with intelligent monitoring generate rich telemetry data. AI agents can predict component failure before it happens:

Predictive Maintenance Agent
  ✓ NeuCool: read pump vibration, flow rate, pressure telemetry
  ✓ NeuCool: read historical component performance baselines
  ✓ Maintenance: create service tickets
  ✓ Alerts: notify NOC and NeuGuard service partner
  ✓ Capacity: pre-stage backup cooling capacity
  × NeuCool: shut down components
  × NeuCool: modify cooling parameters
  × Maintenance: dispatch technicians (requires human approval)

The agent detects that Pump A in CDU-14 shows vibration patterns consistent with bearing wear — 72-hour predicted failure window. It creates a service ticket through NeuGuard’s Authorized Service Partner network, pre-stages cooling redundancy on adjacent CDUs, and alerts the NOC. It cannot shut down the pump itself or dispatch a technician without human approval.

Every prediction, alert, and response is logged. When the NeuGuard warranty team reviews the incident, the audit trail shows exactly what happened, when the degradation was detected, what response was taken, and whether the governed workflow prevented downtime.

5. Multi-Site Infrastructure Governance

Accelsius deploys across hyperscalers, colos, and enterprise data centers — each with different operational policies, SLAs, and risk tolerances. WorkingAgents’ Virtual MCP Servers scope governance per site:

Hyperscaler Site (Equinix Ashburn)
  ✓ Autonomous cooling optimization (full agent authority)
  ✓ Predictive maintenance (auto-create tickets)
  ✓ Workload-aware thermal routing
  Guardrails: thermal limits per Equinix SLA, power caps per contract

Edge Deployment (Remote, Unmanned)
  ✓ Autonomous cooling with enhanced auto-failover
  ✓ Remote monitoring and telemetry
  ✓ Emergency shutdown authority (agent can shut down cooling if leak detected)
  Guardrails: conservative thermal bounds, immediate NOC notification

Enterprise On-Prem
  ✓ Monitoring and alerting only
  × No autonomous parameter changes (human-approval required for all changes)
  Guardrails: strictest bounds, full audit trail for compliance

Same NeuCool hardware. Same WorkingAgents governance platform. Different policies per site based on operational maturity, staffing model, and risk tolerance.

6. Sustainability Reporting and ESG Compliance

Accelsius claims 50% energy reduction vs. air cooling and PUE of 1.08. These claims need continuous measurement, reporting, and verification for ESG reporting and sustainability commitments. WorkingAgents provides:

When a hyperscaler reports “our AI workloads run at PUE 1.08 using Accelsius NeuCool,” the WorkingAgents audit trail proves it — rack by rack, hour by hour, job by job.

Partnership Model

Phase 1: Telemetry Integration (Weeks 1-6)

Phase 2: Joint Pilot (Weeks 7-14)

Phase 3: Autonomous Operations Product (Weeks 15-24)

Revenue Opportunity

The AI data center cooling market is growing explosively:

WorkingAgents monetizes the governance layer on top of Accelsius’s hardware. Accelsius sells NeuCool with governed autonomous management as a differentiator over competitors (CoolIT, GRC, Motivair) who sell hardware without an intelligence layer. The joint offering commands a premium because it delivers measurable outcomes: lower PUE, predictive maintenance, cost attribution, and audit-ready sustainability reporting.

Why This Partnership Works

AI infrastructure is entering a phase where the physical and digital layers can no longer be managed independently. A 250 kW rack running NVIDIA B200s generates enough heat to warm a house. The cooling system keeping those GPUs alive is as mission-critical as the GPUs themselves. And as cooling systems become intelligent and autonomous, they need the same governance that any autonomous system demands: permissions, guardrails, audit trails, and human oversight for high-risk decisions.

Accelsius has built the best thermal solution for AI-scale computing — 4,500W per socket, PUE 1.08, two-phase efficiency that single-phase can’t match. WorkingAgents has built the governance platform for autonomous AI systems — permissions, guardrails, and audit trails at every decision point.

Accelsius removes the thermal barrier to AI scale. WorkingAgents removes the trust barrier to autonomous infrastructure. Together, they deliver what the next generation of AI data centers needs: cooling that’s intelligent, autonomous, and governed — from the chip to the audit trail.

James Aspinwall is the founder of WorkingAgents, an AI governance platform specializing in agent access control, security, and integration services for enterprises deploying AI at scale.