Google introduce new Physical-Engineering Exam for agents & forecast it’s the last iteration

Published: November 5, 2026

Executive Overview

The trajectory of foundation model benchmarking has reached an acute inflection point. While contemporary frontier models routinely achieve saturation scores exceeding 90% across standardized academic examinations, code synthesis benchmarks, and synthetic reasoning batteries (such as MMLU, GSM8K, and HumanEval), these metrics exhibit an alarming decoupling from real-world professional competence. This evaluation gap is nowhere more pronounced than in physical engineering—the disciplined practice of civil, mechanical, structural, and electrical engineering that designs and maintains skyscrapers, transportation grids, energy networks, and semiconductor fabs. In these high-consequence domains, failure is not measured in software stack traces or minor user experience degradation; it carries the potential for catastrophic structural collapse, grid-level blackouts, severe financial liability, and loss of human life.

Historically, the AI research ecosystem has evaluated model reasoning through narrow, clean, and self-contained abstractions: solving textbook math equations, answering multi-choice knowledge questions, or executing isolated Python scripts against synthetic unit tests. Even emerging professional benchmarks—such as Humanity’s Last Exam (HLE) evaluating general multi-disciplinary knowledge or Agents’ Last Exam (ALE) measuring digital enterprise tasks—remain predominantly confined to the digital sphere. They fundamentally fail to test whether an autonomous agent can ingest messy, multi-format project folders, interpret governing municipal building codes, adhere to physical governing equations, balance material cost efficiency, and deliver production-ready engineering specifications capable of withstanding the scrutiny of a licensed Professional Engineer (PE).

To close this evaluation void, the Google AI & Infrastructure team—led by Burak Gokturk (VP Research & Development), Rujun Han (Research Scientist), and George Lee (Product Manager)—in collaboration with Mercor and licensed practicing engineers from industry, unveiled Physical-Engineering’s Last Exam (PLE) on November 3, 2026. Released alongside an open competition leaderboard at ple-bench.com, an open-source evaluation container, and 29 public calibration tasks, PLE establishes the industry’s first rigorous, end-to-end benchmark for evaluating whether autonomous AI agents can execute long-horizon physical engineering work.

Comprising 85 realistic, long-horizon tasks spanning 5,170 engineering files, PLE evaluates agents inside containerized project workspaces (built on Mercor’s Archipelago platform) that faithfully replicate real-world engineering handoffs: PDFs, 2D CAD blueprints, 3D Building Information Modeling (BIM) files, legacy spreadsheets, code standards, and email change-orders. The empirical baseline results published by Google Cloud deliver a sobering reality check to enterprise leaders: across 14 frontier foundation models evaluated over seven distinct agent harnesses, every single model scored below 50%, with the top-performing configuration achieving a modest 48.3%. While frontier models can satisfy individual theoretical constraints in isolation (reaching up to 91% compliance), only 41.3% of their deliverables passed every hard safety and feasibility gate required for real-world construction.

From an IT industry analysis viewpoint, PLE marks a foundational paradigm shift in AI evaluation. As capital expenditure for global physical infrastructure outpaces financing and engineering labor capacity, autonomous agents represent a transformative force multiplier. However, PLE mathematically proves that raw natural-language fluency and software coding prowess do not equate to physical systems reasoning. By introducing hard safety gates, economic cost optimization metrics, and multi-format context reasoning, Google Cloud and Mercor have established the definitive yardstick for Physical AI—providing enterprise engineering firms, infrastructure asset owners, and cloud hyperscalers with an objective standard to measure agent readiness before unvetted models touch the built world.

Features

Physical-Engineering’s Last Exam introduces an advanced evaluation methodology engineered to replicate the operational friction, multimodal context trees, and regulatory constraints of professional engineering practice. The technical architecture comprises several core innovations across environment orchestration, task curation, evaluation mechanics, and benchmark accessibility.

• Long-Horizon Containerized Workspaces on Mercor’s Archipelago Rather than presenting an agent with a single text prompt or an isolated API endpoint, PLE instantiates each task inside a containerized Linux environment powered by Mercor’s Archipelago platform:

  • Realistic File Handoffs: Each task folder contains an average of 61 individual files (with 5,170 total files distributed across the benchmark). These artifacts mirror the heterogeneous file types practicing engineers handle daily: complex PDF specifications, municipal zoning amendments, scanned legacy engineering calculations, Microsoft Excel calculation workbooks, 2D Computer-Aided Design (CAD) drawings (.dwg, .dxf), and 3D Building Information Modeling (BIM) architectures (.ifc, .rvt).
  • Asynchronous Change Orders and Multi-Format Context: Tasks embed realistic human communication artifacts, including email threads containing conflicting client requirements, vendor equipment cut-sheets, and regulatory revision notices. Agents must autonomously parse, cross-reference, and reconcile these documents to isolate the governing physical limits and applicable statutory codes.
  • Native Tool Execution: Inside the Archipelago container, agents have access to standard command-line environments, numerical solvers, Python scientific computing stacks (NumPy, SciPy, SymPy), and file parsers, requiring the agent to synthesize, test, and debug its own numerical calculations.

• Comprehensive Four-Discipline Engineering Taxonomy PLE spans 85 meticulously constructed long-horizon tasks distributed across four foundational engineering disciplines:

  • Mechanical Engineering (30 Tasks): Encompasses thermodynamic cycles, HVAC heating and cooling load calculations under ASHRAE standards, industrial piping hydraulic networks, pump sizing, pressure vessel design under ASME Section VIII codes, and thermal dissipation systems.
  • Electrical Engineering (27 Tasks): Focuses on commercial power distribution, National Electrical Code (NEC) ampacity and conduit sizing, fault current calculations, protective relay coordination curves, transformer loading, and motor control center schematics.
  • Civil Engineering (18 Tasks): Covers stormwater runoff hydrology calculations, open-channel hydraulics, geotechnical soil mechanics and foundation bearing capacity, municipal zoning setbacks, and roadway grading profiles.
  • Structural Engineering (10 Tasks): Evaluates structural load-path tracing, dead/live/wind/seismic load combinations under ASCE 7-22, reinforced concrete beam and slab sizing under ACI 318, and structural steel connection design under AISC specifications.

• Dual-Track Professional Task Authoring Methodology To ensure the benchmark tests authentic professional practice rather than artificial academic puzzles, the 85 tasks were constructed through two distinct methodologies involving licensed practicing engineers:

  • Expert-Authored Tasks (44 Tasks): Designed and built from scratch by practicing licensed engineers (PEs). These experts assembled the raw project folders, drafted ambiguous client problem statements, curated governing code excerpts, and authored deterministic, programmatically executed grading rubrics based on real-world engineering assignments.
  • Co-Authored Tasks (41 Tasks): Practicing engineers designed the initial core task specifications. An automated agent pipeline subsequently assembled the containerized environment and synthetic grading scripts derived from licensable public and private engineering standards. Licensed engineers then thoroughly audited, validated, and increased the difficulty of every task to eliminate edge-case ambiguities and ensure absolute physical correctness.

• Strict Two-Stage Grading Harness: Hard Safety Gates and Economic Optimization PLE introduces an uncompromising, two-stage evaluation rubric that mirrors professional engineering review:

  • Stage 1: Zero-Tolerance Hard Safety and Feasibility Gates: In physical engineering, a design that is 95% compliant is a catastrophic failure if the remaining 5% violates structural deflection limits, electrical thermal thresholds, or hydraulic pressure ratings. Under PLE, an agent receives an absolute score of zero if its deliverable breaches any single hard safety constraint, statutory code limit, or physical conservation law.
  • Stage 2: Economic Cost and Material Optimization: If and only if the design clears 100% of the hard safety gates, the grading harness calculates the agent’s score based on economic optimization—measuring capital expenditure (CapEx), material efficiency, energy consumption, and lifecycle operational cost against established industry engineering baselines.

• Extensive Baseline Benchmarking Matrix (14 Models x 7 Harnesses) Google Cloud and Mercor subjected 14 contemporary frontier models to rigorous benchmarking across seven distinct, publicly available open-source and commercial agent orchestration harnesses:

  • Evaluated Harnesses: Antigravity, Claude Code, Codex, Hermes, OpenCode, OpenHands, and Pi.
  • Disproportionate Model Influence: The empirical findings reveal that underlying foundation model intelligence dictates performance far more than harness design: the benchmark recorded up to a 50-point spread across different models on the same harness, compared to a median 3-point spread across different harnesses utilizing the identical underlying model. However, harness selection exerted a massive influence on operational execution cost and token consumption.

• Open Science Architecture and Competition Framework To catalyze rapid research iteration, Google and Mercor published a comprehensive open-evaluation package:

  • Public Calibration Dataset: 29 complete tasks, grading scripts, and the Docker evaluation container are publicly accessible on GitHub.
  • Private Evaluation Leaderboard (ple-bench.com): Teams can submit custom agent pipelines and proprietary architectures to be evaluated against the remaining 56 private holdout tasks.
  • Subsidized Compute: Google Cloud provides free, serverless hosting and evaluation infrastructure for 13 Google-sponsored models on the competition platform, lowering barrier-to-entry for academic and independent research groups.

Benefits

The introduction of Physical-Engineering’s Last Exam delivers transformative strategic, technical, and operational advantages to enterprise engineering firms, infrastructure asset operators, foundation model researchers, and cloud platform providers.

• Elimination of “Evaluation Illusion” for High-Consequence Engineering AI For enterprise technology executives, one of the greatest hazards in deploying generative AI has been the “evaluation illusion”—the mistaken belief that high scores on software coding or academic benchmarks indicate readiness for physical systems design. A model scoring 95% on HumanEval or MMLU can easily hallucinate pipe diameters, misapply seismic load factors, or invert thermal conductivity constants. PLE provides the global engineering community with an unvarnished, empirical baseline. Enterprise decision-makers can now quantitatively separate speculative marketing claims from proven physical reasoning capability, ensuring that AI tools are integrated strictly within realistic operational boundaries.

  
  

• Definitive Acceleration Toward Closing the Global Infrastructure Deficit Across global economies, the demand for infrastructure modernization—spanning clean energy transition grids, advanced water reclamation systems, transportation networks, and next-generation data center construction—is severely outpacing available financing and skilled engineering labor. The engineering talent pool is constrained by an aging workforce and lengthy professional licensing pipelines. By establishing a rigorous benchmark that measures agents against end-to-end engineering deliverables, PLE provides foundation model labs with a clear optimization roadmap. Advancing agent scores on PLE directly translates into developing reliable digital assistants capable of automating preliminary calculations, code compliance sweeps, and drafting workflows, effectively multiplying the productivity of licensed engineers and compressing project delivery timelines.

  
  

• Establishment of “Safety-First” Objective Optimization Functions In conventional software agent evaluation, accuracy is graded on continuous or soft curves (e.g., partial credit for test cases passed). PLE fundamentally rejects this paradigm by establishing binary, zero-tolerance hard safety gates. By penalizing safety violations with an absolute score of zero, PLE forces the AI research community to align loss functions and reinforcement learning with verifiable physical constraints. Models optimized against PLE learn that hallucinating or cutting corners on safety codes completely invalidates an answer, instilling rigorous conservatism into AI design loops.

  
  

• Standardization of Multimodal and Multi-Document Engineering Reasoning Real-world engineering problems never present themselves as clean natural-language prompts; they reside across dozens of messy, disjointed files. By requiring agents to navigate 61 files on average per task—integrating raster PDFs, tabular spreadsheets, vector CAD geometries, and email revisions—PLE pushes agent architectures to master complex, long-context retrieval, cross-modal synchronization, and state tracking. This drives architectural innovation in agent memory compaction, document grounding, and tool-calling precision.

  
  

• Empirical FinOps Transparency Across Agent Harnesses The baseline findings from PLE provide crucial economic guidance for platform engineering teams architecting enterprise agent systems. By validating that foundation model choice drives performance (up to a 50-point spread) while agent harnesses primarily influence operational execution cost (a median 3-point performance spread but massive variance in token consumption), PLE demonstrates that optimizing agent harnesses should focus on token efficiency, context compression, and caching, while cognitive performance must be driven by upgrading underlying foundation models.

  
  

Use Cases

The structural methodology, containerized environments, and evaluation metrics established by PLE mirror several mission-critical enterprise deployment patterns where autonomous physical-engineering agents can assist licensed professionals.

• Automated Building Information Modeling (BIM) Code Compliance and Clash Detection Architecture, Engineering, and Construction (AEC) conglomerates manage massive multi-disciplinary BIM files for commercial high-rises and healthcare facilities, requiring weeks of manual cross-referencing against local building codes.

  • Operational Context: An engineering firm is designing a 40-story commercial tower subject to complex municipal fire safety codes, ADA accessibility guidelines, and ASHRAE energy conservation standards.
  • Agentic Execution: Built on the architectural principles of PLE, an autonomous engineering agent is deployed into the containerized project directory. The agent parses the architectural Revit/IFC files, extracts floor layouts, and interrogates local building regulations. It autonomously executes egress travel distance calculations, verifies emergency stairwell fire-rating assemblies, and checks mechanical duct clearances against structural steel framing.
  • Business Outcome: The agent generates a comprehensive compliance report highlighting three non-compliant egress corridors and four mechanical-structural clashes, drafting revised CAD geometry that satisfies all governing safety codes for human PE sign-off, compressing preliminary plan review cycles from three weeks to four hours.

• Automated Distribution Substation Sizing and Electrical Load Flow Analysis Electric utility operators face massive backlogs interconnecting distributed solar arrays, battery storage facilities, and high-density AI data center loads to regional power grids.

  • Operational Context: An electrical utility engineering department must evaluate an interconnection request for a 200MW hyperscale data center campus requiring 115kV transmission step-down substations.
  • Agentic Execution: An agent running in a PLE-aligned environment ingests utility line diagrams, transformer manufacturer specifications, and the regional transmission operator’s interconnect criteria. The agent executes short-circuit fault current analyses, models voltage drop under extreme peak loads, sizes underground cable conduits under NEC ampacity derating tables, and calculates protective relay trip curves.
  • Business Outcome: The agent outputs a complete electrical calculation packet, single-line diagrams, and bill-of-materials that satisfy utility reliability criteria while optimizing transformer impedance to minimize capital equipment cost, accelerating interconnect approval timelines without compromising grid stability.

• Industrial Plant Piping Hydraulic Balancing and Pump Optimization Petrochemical, pharmaceutical, and water treatment processing plants operate intricate, closed-loop fluid piping networks governed by strict ASME pressure vessel and process piping codes.

  • Operational Context: A chemical processing facility must expand its cooling water distribution loop to service three new reactor modules without exceeding line pressure limits.
  • Agentic Execution: The agent is provisioned with plant isometric drawings, fluid viscosity profiles, pipe roughness factors, and pump performance curves. Navigating the project folder, the agent applies the Darcy-Weisbach equation to calculate friction head losses across varying flow velocities, models hydraulic transient pressures (water hammer) during emergency shutdown events, and selects optimal centrifugal pump impellers.
  • Business Outcome: Adhering to PLE’s two-stage grading philosophy, the agent ensures that maximum operating pressure remains safely below ASME B31.3 limits before selecting a variable-frequency drive pump configuration that slashes annual operational energy consumption by 18%, saving the plant hundreds of thousands of dollars in lifecycle power costs.

• Geotechnical Foundation Sizing and Stormwater Runoff Hydrology Civil engineering infrastructure projects require detailed analysis of heterogeneous soil profiles, seismic soil liquefaction risks, and municipal stormwater retention mandates.

  • Operational Context: A civil engineering consultancy is engineering the site development for a 50-acre industrial logistics hub requiring retention ponds, foundation grading, and pavement structural design.
  • Agentic Execution: The agent analyzes soil boring logs, geotechnical laboratory test results, and regional NOAA precipitation frequency tables. It executes the Rational Method and SCS runoff curve number equations to model 100-year storm flood volumes, sizes retention basin geometries, and calculates ultimate bearing capacity for shallow spread footings under variable groundwater conditions.
  • Business Outcome: The agent produces stamped-ready civil grading plans and drainage calculations that guarantee zero downstream flood surcharge while optimizing earthwork cut-and-fill balancing, saving the developer significant site-clearing and soil-export costs.

Alternatives

Enterprise engineering leaders, research laboratories, and foundation model developers evaluating Physical-Engineering’s Last Exam must understand how it compares against existing AI evaluation benchmarks and specialized domain verification frameworks.

• Alternative 1: Humanity’s Last Exam (HLE) and MMLU-Pro

  • Architectural Overview: Humanity’s Last Exam and Massive Multitask Language Understanding (MMLU-Pro) represent premier cross-disciplinary knowledge benchmarks. They utilize multiple-choice and short-answer questions designed by global subject-matter experts to test the absolute frontier of academic and scientific knowledge across mathematics, physics, philosophy, and biology.
  • Structural Comparison: While HLE and MMLU-Pro present exceptionally difficult theoretical problems that challenge model reasoning, they fundamentally test static, declarative knowledge. They do not evaluate multi-turn agentic tool calling, execution inside file-system workspaces, reading messy CAD drawings, or producing production-ready engineering documentation. An agent can achieve an elite score on HLE by memorizing theoretical fluid dynamics formulas while remaining completely incapable of sizing a real-world building drainage pipe under municipal building codes. PLE is an operational, task-based benchmark; HLE is an academic knowledge benchmark.

• Alternative 2: Agents’ Last Exam (ALE) and SWE-bench

  • Architectural Overview: Agents’ Last Exam (ALE) and SWE-bench evaluate autonomous agents executing digital professional work—primarily software engineering tasks. SWE-bench evaluates an agent’s ability to resolve genuine GitHub issues by cloning a repository, reading source code, synthesizing patches, and passing PyTest unit tests inside a Docker container.
  • Structural Comparison: SWE-bench pioneered the containerized, multi-file agentic evaluation methodology that PLE adopts. However, software engineering operates within a deterministic digital environment where syntax errors throw immediate compiler feedback, and success is validated by binary software unit tests. Physical engineering introduces governing physics, statutory municipal codes, and economic optimization trade-offs that do not exist in pure software development. A software patch can be refactored endlessly in production CI/CD; a poured concrete foundation cannot. PLE expands the containerized agentic paradigm into the physical and legal constraints of the built world.

• Alternative 3: Commercial Physics Solvers and CAE Suites (ANSYS, COMSOL, Autodesk Revit, Bentley)

  • Architectural Overview: Traditional engineering design relies on Computer-Aided Engineering (CAE) and Finite Element Analysis (FEA) platforms. These suites utilize deterministic numerical solvers to simulate structural stress, fluid dynamics, and electromagnetic fields based on explicit mathematical models inputted by human engineers.
  • Structural Comparison: CAE software suites are deterministic calculation tools, not cognitive evaluation benchmarks. They do not formulate design concepts, interpret ambiguous human requests, or cross-reference municipal code documents. PLE does not seek to replace ANSYS or Revit; rather, it benchmarks whether autonomous AI agents can intelligently orchestrate tools, drive engineering calculations, interpret code standards, and utilize software stacks to generate valid engineering designs.

• Alternative 4: Academic Engineering Exams (FE and PE Practice Examinations)

  • Architectural Overview: The National Council of Examiners for Engineering and Surveying (NCEES) administers the Fundamentals of Engineering (FE) and Principles and Practice of Engineering (PE) exams—standardized examinations required for professional licensure in the United States.
  • Structural Comparison: AI developers frequently test models against digitized FE and PE practice exam questions. However, these exams are designed to test individual human cognitive retention under timed examination conditions. They consist of isolated, multiple-choice questions with simplified boundary conditions. Real-world engineering projects are never presented as multiple-choice questions; they involve messy, uncurated project folders with incomplete information, conflicting emails, and complex CAD files. PLE tests the actual job of an engineer, not their ability to pass a standardized classroom test.

An Alternative Perspective

A rigorous, objective industry critique must look beyond the excitement surrounding Google Cloud and Mercor’s announcement to critically examine the structural limitations, operational assumptions, and potential risks inherent in Physical-Engineering’s Last Exam. While PLE represents an essential advancement in AI evaluation, an alternative perspective reveals profound challenges that enterprise engineering leaders must carefully navigate.

First, the benchmark’s sub-50% baseline score across all frontier models highlights the immense, unbridged chasm between current LLM architectures and autonomous physical systems deployment. The fact that the highest-performing model configuration achieved only 48.3%, and that only 41.3% of deliverables passed all hard safety checks, demonstrates that today’s frontier models are fundamentally unsafe to operate autonomously in physical engineering. If an enterprise engineering firm were to deploy current agents to generate construction-ready documents without exhaustive, line-by-line human oversight, the risk of structural, electrical, or hydraulic failure would be catastrophic. The primary takeaway of PLE is not that AI is ready to transform engineering, but that current foundation models remain deeply unreliable in physical systems reasoning.

Second, the structural reliance on synthetic and co-authored tasks introduces potential benchmark contamination and verification fragility. Google Cloud notes that 41 of the 85 tasks were “co-authored,” wherein an automated agent pipeline assembled the task environment and grading scripts from licensable standards, which human engineers then validated. In complex engineering domains, subtle edge-case interactions between different sections of engineering codes (e.g., how seismic design categories modify wind load calculations in ASCE 7) frequently produce contradictory interpretations even among experienced licensed engineers. If the automated agent pipeline that constructed the grading rubrics embedded subtle misinterpretations of the codes, the benchmark could unfairly penalize valid engineering designs that adopted alternative, equally valid professional methodologies, or conversely, award points to dangerous shortcuts that human validators overlooked.

Third, the containerized Archipelago environment—while impressive—cannot fully capture the iterative, sociotechnical reality of professional engineering. In real-world engineering practice, designs are not created in a vacuum inside an isolated file folder. Engineering is fundamentally a collaborative, conversational negotiation: practicing engineers conduct physical site walks to inspect field conditions, interview facility managers to understand operational nuances, negotiate variances with municipal building inspectors, and engage in continuous back-and-forth reviews with client stakeholders. Reducing professional engineering to a static set of 61 files and a deterministic grading script risks over-indexing on mathematical optimization while completely neglecting the vital human communication, ethical judgment, and physical intuition that define professional engineering responsibility.

Fourth, the benchmark’s geographic and regulatory bias limits its global applicability. Physical engineering codes are heavily regionalized: the United States relies on IBC, NEC, and ASCE; Europe enforces Eurocodes; and jurisdictions across Asia maintain distinct statutory building and seismic regulations. Because PLE is heavily anchored in United States building codes and professional standards (such as ASHRAE, ASME, ACI, and NEC), a model optimized against PLE may develop acute American regulatory bias. Deploying such a model in London, Tokyo, or Singapore could result in severe regulatory non-compliance, as the agent may attempt to apply American National Electrical Code conduit sizing to European electrical topologies or British structural steel standards.

Finally, the legal liability and professional licensure frameworks governing physical engineering present an absolute structural barrier to autonomous agent adoption that no benchmark can resolve. In every developed jurisdiction, engineering designs submitted for construction permitting must be stamped and sealed by a licensed Professional Engineer (PE). Licensure laws legally mandate that the signing engineer must have exercised “direct supervisory control” over the calculations and drawings. If a building collapses, a software vendor cannot be prosecuted for professional malpractice; the human engineer who stamped the drawings bears full civil, and potentially criminal, liability. Therefore, even if a future foundation model achieves a 99% score on PLE, licensed engineers will remain legally obligated to manually verify every single calculation, dimension, and code reference. Until legal and insurance frameworks evolve to define AI liability in engineering, the return on investment of engineering agents will be bounded by the time required for human engineers to audit the agent’s work.

Final Thoughts

The launch of Physical-Engineering’s Last Exam (PLE) by Google AI & Infrastructure, Mercor, and practicing licensed engineers marks a monumental, long-overdue maturation in artificial intelligence evaluation. By forcing autonomous agents to step out of the synthetic digital sandbox and confront the unforgiving laws of physics, statutory building codes, and material economics, PLE dismantles the pervasive illusion that high academic benchmark scores equate to professional competence. The sobering reality exposed by PLE—that every frontier foundation model currently scores below 50%, with less than 42% passing all hard safety gates—provides the global engineering community with an indispensable, mathematically rigorous reality check.

For Chief Information Officers, engineering practice leaders, and infrastructure platform directors, PLE provides a clear operational framework. Technology leaders must firmly reject the temptation to treat current generative AI as autonomous design engines. Instead, engineering enterprises should utilize PLE’s open-source tasks and evaluation methodology to build controlled, human-in-the-loop digital apprenticeships: deploying agents to automate document ingestion, draft preliminary calculation schedules, and perform initial code-compliance audits, while reserving all design authorization and safety validation strictly for licensed Professional Engineers.

As the AI research community converges on the open competition at ple-bench.com, the benchmark will undoubtedly serve as the primary catalyst for the next era of foundation model development: transitioning models from probabilistic natural-language mimics into mathematically grounded, physically consistent systems of action. The path to building the physical world with artificial intelligence requires uncompromising adherence to structural integrity, public safety, and economic prudence; with Physical-Engineering’s Last Exam, the industry finally has the ruler to measure that journey.

Source

The source blog post utilized for this analysis can be accessed at: Introducing Physical-Engineering’s Last Exam for agents | Google Cloud Blog