Phaidra

Agents read from and write to named real systems. The Phaidra Factory page describes the PUE agent managing the chiller plant (chiller staging, evaporator temperatures, differential pressures) "via a BMS/SCADA integration"; the CDU agent "preemptively controls the CDU"; and Agentic Power Allocation "integrates directly with NVIDIA's Mission Control Domain Power Service (DPS) via the NVGrid API" to change compute allocation, and updates power allocation policies using NVIDIA DSX Max-Q APIs. The Merck case study adds that the cloud service "updates the local BMS directly".

Phaidra is also a named contributor to the NVIDIA DSX and Omniverse digital twin blueprints, which is a partnership rather than an integration.

Credential scoping and custom tools are not documented.

Sourcephaidra.ai/products/phaidra-factory, phaidra.ai/blog Merck case study (25 June 2024); readread 2026-09-15

Coordination between agents runs under a coordinating agent. The Merck case study describes four AI Virtual Plant Operators, one per chiller plant, and a fifth, the "AI Conductor", that takes in campus-level data and "coordinates each individual agent's production of chilled water to optimize load balance between all plants", live since early 2023. That is routing work across autonomous agents under a coordinator.

Deterministic and autonomous steps mix: the agents write setpoints into the existing BMS, which keeps its hard-coded sequence of operations, and a bump-less transfer hands control back to local logic. An agent sending setpoint commands directly to a coolant distribution unit is autonomous execution rather than orchestration; the coordination comes from the AI Conductor.

No authoring surface, versioning or reuse across teams is documented, and the coordinator was built for one campus.

Sourcephaidra.ai/blog Merck case study (25 June 2024), phaidra.ai/products/phaidra-factory; readread 2026-09-15

Answers in Phaidra Prism are grounded in live data, historical trends and system documentation pulled at question time. Its product page describes an assistant, Alfred, that is there "to pull from live data, historical trends and system documentation to respond to questions, provide context and generate charts", and says Prism "provides consistent analysis of historical and real-time data".

Because documentation and operating history are pulled when a question is asked rather than trained in, new customer knowledge can be added without retraining. The retrieval is described in one product paragraph, which does not say which sources are indexed, refreshed and permissioned or whether answers cite their sources.

The control agents ground differently. A controller computing setpoints from live sensor tags is inference on inputs, not querying knowledge, and the agents' own grounding is by training: in the Merck case study, "the AI agent was trained on the plant's historical data".

Sourcephaidra.ai/products/phaidra-prism, phaidra.ai/blog Merck case study (25 June 2024); readread 2026-09-15

Three oversight mechanisms ship in the service: pause and resume, an approval mode and structured guardrails. For pause and resume, the service "has a built-in bump-less transfer to local control triggered by operators at any time", and the agents resume correctly after AI Off periods.

Before autonomous control at Merck, the agent ran in a manual approval mode that "sent setpoint recommendations to on-site staff for their review and implementation", with feedback on rejected changes, and the move to autonomous control was a customer decision.

As for guardrails, the agent was "programmed to understand and respect all system constraints", and on the Factory page Agentic Power Allocation keeps a conservative margin while "maintaining rigorous site-safety guardrails". The transfer, the recommendation mode and the constraints are Phaidra's own mechanisms in the service, not obligations placed on the customer.

Routing approvals into existing ticketing channels, or explaining why a checkpoint fired, is not documented.

Sourcephaidra.ai/blog Merck case study (25 June 2024), phaidra.ai/products/phaidra-factory; readread 2026-09-15

The attestations are in place, but no access control that a customer operates is published. Phaidra's Security page, linked from every phaidra.ai footer, states that Phaidra "has completed SOC 2 Type 2 auditing and attestation", with the report available to customers on request, and "has obtained ISO/IEC 27001:2013 certification" through ANAB.

Access control is described only for Phaidra's own staff: least privilege and role-based access "enforced via the Google IAM service" govern access to Phaidra's production environment, "only specific employees with authorized credentials can access your data", and "all platform access is audit-trailed". Those are internal controls, not an access surface the customer operates. Encryption at rest and in transit with GCP KMS is also documented.

Sourcephaidra.ai/security; readread 2026-09-15

Visibility into what the agents are doing, have done and plan to do is described, but no surface where a customer inspects it is named. Under "AI transparency & explainability", the Security page says Phaidra has developed ways "to illustrate not only what the AI is currently doing, but also how it's performed historically and what it's planning into the future", to turn the black box into a white box for plant operators. That is visibility into the agent's own actions, history and plans.

The claim names no run history, action log, trace view or audit trail of agent decisions that a customer can open, so step by step inspection of inputs, actions and outputs is not documented, nor are audit logs separate from runtime traces. "All platform access is audit-trailed" describes staff access to Phaidra's production environment, not the agents, and Phaidra Prism's charts of chillers, alarms and consumption read the customer's equipment, not the agents.

Sourcephaidra.ai/security, phaidra.ai/products/phaidra-prism; readread 2026-09-15

No agent memory that a customer can see or govern is documented. No memory types (session, conversation, workflow, long-term) and no way to review, edit, delete or scope memories by user, team or workspace appear anywhere on Phaidra's site.

What Phaidra's agents do is learn: an agent trained in simulation improves further after several hours of live learning, and the reinforcement learning agent teaches itself to get better without human intervention. Learning that improves a controller's policy is the model changing, not memory the agent keeps and a customer can govern.

The Merck case study notes that during AI Off times the agents "still receive live operational data", so on reactivation they understand what changed. That is continuous telemetry ingestion while control is paused, not a memory store.

Sourcephaidra.ai/blog Merck case study (25 June 2024); readread 2026-09-15

Deployment comes in one mode, a cloud service, with no region or residency selection. The Merck case study describes "Phaidra's cloud-based service" that analyzes live operational data and updates the local BMS, and the Security page places the service on Google Cloud Platform, with GCP Key Management Service and Google IAM. No on-premises, private cloud, VPC or edge option is documented, no region is named, and nothing lets a customer pin model traffic or data storage to specific regions or infrastructure.

Writing setpoints into on-site BMS and SCADA is integration, not deployment topology, and GCP as a hosting provider is infrastructure, not a deployment option.

Sourcephaidra.ai/security, phaidra.ai/blog Merck case study; readread 2026-09-15

Agent solutions are named for buyers but delivered as per facility deployments rather than a catalog to adopt.

Phaidra sells two named products, Phaidra Prism and Phaidra Factory, and the Factory page presents three named agent solutions, each with a stated challenge, solution and result: CDU control (the Liquid Cooling Agent), PUE optimization (a chiller plant agent via BMS/SCADA) and increasing IT capacity (Agentic Power Allocation).

The page calls these "just two of several Phaidra agents", with "more agents to be released this year". These are buyer-facing solutions, not internal machinery.

The three solutions share one page with no individual entries, and each is built for a particular facility rather than adopted ready made: the Merck case study describes nearly a year of joint work to "develop and deploy" an agent per chiller plant, and the control agents are trained on each plant's history. These are production deployments, not starting points.

Sourcephaidra.ai/products/phaidra-factory, phaidra.ai/applications, phaidra.ai/blog Merck case study; readread 2026-09-15

Work reaches Phaidra's agents without a person starting it, on two documented autonomous seams. Event-driven: the Phaidra Factory page says the Liquid Cooling Agent "monitors GPUs to predict thermal spikes" and, "when conditions suggest a spike is imminent", begins the cooling process, with response times under 10 seconds, and the PUE agent proactively manages the chiller plant.

Continuous: the Merck case study describes a cloud service that analyzes live operational data and updates the local BMS directly, with agents still receiving live data during AI Off periods so they respond correctly on reactivation. Phaidra Prism also alerts operators to outliers and anomalies from historical and real-time data without being asked. In each case the agent starts on the signal, not when asked.

The seams are telemetry and anomaly events only; no schedule, webhook or messaging channel is documented.

Sourcephaidra.ai/products/phaidra-factory, phaidra.ai/products/phaidra-prism, phaidra.ai/blog Merck case study; readread 2026-09-15

Model choice on both Phaidra products is vendor-controlled. The control agents are Phaidra's own reinforcement learning models trained per facility, per the Merck case study. Phaidra Prism is described as "the only LLM designed for data center operators", but no underlying model or provider is named, no routing across models is disclosed, and no customer selection or bring-your-own-key option exists. GCP as the hosting provider is infrastructure, not a model.

Sourcephaidra.ai/products/phaidra-prism, phaidra.ai/blog Merck case study; readread 2026-09-15

No API, SDK or MCP surface for the platform is published, so the product is not callable or composable from outside.

Agentic Power Allocation calls NVIDIA's NVGrid and DSX Max-Q APIs, but that is Phaidra consuming another vendor's API, not a surface that makes Phaidra callable.

Sourcephaidra.ai/products/phaidra-factory, phaidra.ai/applications; readread 2026-09-15

Phaidra ran a controlled trial with a readable, comparable result before any autonomous control was allowed.

The Merck case study documents a month-long demonstration in which the agent's recommendations ran against the plant's existing sequence of operations, with a published comparative performance analysis (Figure 1): 16.2 percent energy savings against the local SOO, 70.5 percent better thermal stability and 50.9 percent less excess equipment runtime, after which the customer approved the change.

That is a change under test before release, with a result readable by the customer who approved it. After deployment, staff feedback on recommendations they did not implement was used to improve the agent.

Training agents in simulation before deployment is how the model is built, not agent evaluation.

No standing evaluation surface, test fixtures or datasets a customer runs are documented; the trial is run within Phaidra's deployment process.

Sourcephaidra.ai/blog Merck case study (25 June 2024); readread 2026-09-15

Every action Phaidra documents goes through a machine interface, so browser and computer control has no role in the product. The PUE agent works through a BMS/SCADA integration, the Liquid Cooling Agent controls the CDU, Agentic Power Allocation uses NVIDIA's NVGrid API, and in the Merck deployment the service updates the local BMS directly. Phaidra Prism generates charts for operators, which is output, not computer control.

Sourcephaidra.ai/products/phaidra-factory, phaidra.ai/blog Merck case study; readread 2026-09-15