{"version":1,"accessed_at":"2026-09-02","sources":[{"id":"mts-trace","publisher":"MTS Intelligence","title":"Representative duration-fix trace, raw observations, and capture methodology","url":"/api/agent-infrastructure/traces/default","published_at":"2026-09-02","accessed_at":"2026-09-02","evidence_type":"mts-measurement","scope":"Local environment-side timings, outcomes, process launches, peak RSS, and bounded logical fixture bytes.","qualification":"No live model was invoked; provider telemetry, total I/O, utilization, energy, and allocation are missing."},{"id":"openai-agent-loop","publisher":"OpenAI","title":"Unrolling the Codex agent loop","url":"https://openai.com/index/unrolling-the-codex-agent-loop/","published_at":"2026-01-23","accessed_at":"2026-09-02","evidence_type":"primary-documentation","scope":"Turn construction, repeated model/tool interactions, context, and tool results.","qualification":"Public conceptual documentation, not task-level production telemetry."},{"id":"openai-harness","publisher":"OpenAI","title":"Unlocking the Codex harness: how we built the App Server","url":"https://openai.com/index/unlocking-the-codex-harness/","published_at":"2026-02-04","accessed_at":"2026-09-02","evidence_type":"primary-documentation","scope":"Harness state, protocol, approvals, tool execution, and event processing.","qualification":"Documents Codex components without exposing provider hardware measurements."},{"id":"openai-safety","publisher":"OpenAI","title":"Running Codex safely at OpenAI","url":"https://openai.com/index/running-codex-safely/","published_at":"2026-05-08","accessed_at":"2026-09-02","evidence_type":"primary-documentation","scope":"Execution isolation and security controls.","qualification":"A security architecture source, not a performance benchmark."},{"id":"anthropic-managed-agents","publisher":"Anthropic","title":"Scaling Managed Agents: Decoupling the brain from the hands","url":"https://www.anthropic.com/engineering/managed-agents","published_at":"2026-04-08","accessed_at":"2026-09-02","evidence_type":"primary-documentation","scope":"Remote agent orchestration and the separation between model reasoning and execution environments.","qualification":"First-party implementation account; do not generalize to all agents."},{"id":"anthropic-sandboxing","publisher":"Anthropic","title":"Beyond permission prompts: making Claude Code more secure and autonomous","url":"https://www.anthropic.com/engineering/claude-code-sandboxing","published_at":"2025-10-20","accessed_at":"2026-09-02","evidence_type":"primary-documentation","scope":"Filesystem and network sandbox boundaries for local tool execution.","qualification":"Describes Claude Code mechanisms, not universal container use."},{"id":"anthropic-containment","publisher":"Anthropic","title":"How we contain Claude across products","url":"https://www.anthropic.com/engineering/how-we-contain-claude","published_at":"2026-05-25","accessed_at":"2026-09-02","evidence_type":"primary-documentation","scope":"OS policies, containers, VMs, and product-dependent isolation boundaries.","qualification":"Containment mechanisms vary by product and deployment."},{"id":"anthropic-evals","publisher":"Anthropic","title":"Demystifying evals for AI agents","url":"https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents","published_at":"2026-01-09","accessed_at":"2026-09-02","evidence_type":"primary-documentation","scope":"Agent evaluation design and the distinction between outcomes and trajectories.","qualification":"Evaluation guidance, not hardware telemetry."},{"id":"codex-core","publisher":"OpenAI","title":"codex-core README","url":"https://github.com/openai/codex/blob/a2a9a434764b9ff875022c028f636a77656d1abc/codex-rs/core/README.md#L21-L63","published_at":"2026-09-02","accessed_at":"2026-09-02","evidence_type":"source-code","scope":"Inspectable turn-loop and session implementation notes.","qualification":"Commit-pinned source; current behavior may evolve."},{"id":"codex-exec","publisher":"OpenAI","title":"Codex unified exec command handler","url":"https://github.com/openai/codex/blob/a2a9a434764b9ff875022c028f636a77656d1abc/codex-rs/core/src/tools/handlers/unified_exec/exec_command.rs#L131-L132","published_at":"2026-09-02","accessed_at":"2026-09-02","evidence_type":"source-code","scope":"Inspectable process-execution boundary.","qualification":"A code pointer, not a measured production workload."},{"id":"mcp-tools","publisher":"Model Context Protocol","title":"MCP specification: Tools","url":"https://modelcontextprotocol.io/specification/2025-06-18/server/tools","published_at":"2025-06-18","accessed_at":"2026-09-02","evidence_type":"primary-documentation","scope":"Tool discovery and invocation semantics.","qualification":"Protocol interface; physical execution topology is implementation-specific."},{"id":"mcp-transports","publisher":"Model Context Protocol","title":"MCP specification: Transports","url":"https://modelcontextprotocol.io/specification/2025-06-18/basic/transports","published_at":"2025-06-18","accessed_at":"2026-09-02","evidence_type":"primary-documentation","scope":"Local stdio and remote HTTP transport alternatives.","qualification":"Transport choice alone does not reveal workload cost."},{"id":"swe-rex","publisher":"SWE-agent","title":"SWE-agent Remote Execution Framework","url":"https://github.com/SWE-agent/SWE-ReX/tree/5c995c365dfb1fd5bc56fda688be5d8538f9931f","published_at":"2026","accessed_at":"2026-09-02","evidence_type":"source-code","scope":"Execution substrate across Docker, Kubernetes, modal, and cloud backends.","qualification":"Runtime infrastructure, not an end-to-end hardware benchmark."},{"id":"swe-bench","publisher":"ICLR","title":"SWE-bench: Can Language Models Resolve Real-World GitHub Issues?","url":"https://proceedings.iclr.cc/paper_files/paper/2024/file/edac78c3e300629acfe6cbe9ca88fb84-Paper-Conference.pdf","published_at":"2024","accessed_at":"2026-09-02","evidence_type":"research-paper","scope":"Task-resolution benchmark for real GitHub issues.","qualification":"Measures whether issues are resolved, not resource-level critical paths."},{"id":"swe-bench-verified","publisher":"OpenAI","title":"Introducing SWE-bench Verified","url":"https://openai.com/index/introducing-swe-bench-verified/","published_at":"2024-08-13","accessed_at":"2026-09-02","evidence_type":"primary-documentation","scope":"Human-validated subset of SWE-bench.","qualification":"Improves task validity without adding infrastructure telemetry."},{"id":"agentperf","publisher":"Artificial Analysis","title":"AgentPerf Methodology","url":"https://artificialanalysis.ai/methodology/agentperf","published_at":"2026","accessed_at":"2026-09-02","evidence_type":"benchmark-repository","scope":"Live inference-platform testing over agentic request patterns.","qualification":"Tool delays are simulated; environment execution is not physically measured."},{"id":"mlperf","publisher":"MLCommons","title":"MLPerf Inference benchmark suite","url":"https://github.com/mlcommons/inference/tree/3fbc329939999c13d0a7b5e67fb2092287e06047","published_at":"2026","accessed_at":"2026-09-02","evidence_type":"benchmark-repository","scope":"Standardized model-inference performance and optional power measurement.","qualification":"A bounded inference query is not a full agent loop."},{"id":"agentsysbench","publisher":"arXiv","title":"From LLM Inference to Agentic Workloads: Characterization and Implications for Serving Systems","url":"https://arxiv.org/abs/2608.15127","published_at":"2026-08-15","accessed_at":"2026-09-02","evidence_type":"research-paper","scope":"Early full-loop characterization of model and non-model latency, memory, and CPU allocation across agent applications.","qualification":"August 2026 preprint; code was not released at review time and the results were not independently reproduced."},{"id":"autellix","publisher":"arXiv","title":"Autellix: An Efficient Serving Engine for LLM Agents as General Programs","url":"https://arxiv.org/abs/2502.13965","published_at":"2025-02-19","accessed_at":"2026-09-02","evidence_type":"research-paper","scope":"Program-level scheduling and LLM-call distributions across ShareGPT, BFCL, and LATS evaluation workloads.","qualification":"Author-run systems evaluation; its call-count distributions demonstrate possible program shapes, not a typical count for coding agents."},{"id":"nvidia-vera","publisher":"NVIDIA","title":"NVIDIA Vera CPU","url":"https://www.nvidia.com/en-us/data-center/vera-cpu/","published_at":"2026","accessed_at":"2026-09-02","evidence_type":"vendor-claim","scope":"Vera architecture, CPU role, memory, and fabric specifications.","qualification":"Vendor specifications and positioning; not independent agent-workload measurement."},{"id":"nvidia-vera-agents","publisher":"NVIDIA","title":"NVIDIA Vera CPU Boosts AI Factory Throughput to Accelerate Agentic Workloads","url":"https://developer.nvidia.com/blog/nvidia-vera-cpu-boosts-ai-factory-throughput-to-accelerate-agentic-workloads/","published_at":"2026-07-07","accessed_at":"2026-09-02","evidence_type":"vendor-claim","scope":"Vendor sandbox-throughput benchmark and agent-workload argument.","qualification":"The 1.8x result uses an unnamed x86 baseline and lacks enough public configuration detail for neutral reproduction."},{"id":"nvidia-olympus","publisher":"NVIDIA","title":"Inside NVIDIA Vera CPU: Olympus Cores Built for Maximum Single-Threaded Performance in Agentic AI","url":"https://developer.nvidia.com/blog/inside-nvidia-vera-cpu-olympus-cores-built-for-maximum-single-threaded-performance-in-agentic-ai","published_at":"2026-07-21","accessed_at":"2026-09-02","evidence_type":"vendor-claim","scope":"Olympus cores, threading, branch behavior, cache, and loaded-core design claims.","qualification":"Architecture is described by its vendor; comparative workload claims are not neutral reproduction."},{"id":"nvidia-rubin","publisher":"NVIDIA","title":"NVIDIA Vera Rubin NVL72","url":"https://www.nvidia.com/en-us/data-center/vera-rubin-nvl72/","published_at":"2026","accessed_at":"2026-09-02","evidence_type":"vendor-claim","scope":"Integrated 72-Rubin-GPU and 36-Vera-CPU rack-scale platform.","qualification":"Preliminary vendor specifications; product configuration may change."},{"id":"nvidia-rubin-platform-blog","publisher":"NVIDIA","title":"Inside the NVIDIA Vera Rubin Platform: Six New Chips, One AI Supercomputer","url":"https://developer.nvidia.com/blog/inside-the-nvidia-rubin-platform-six-new-chips-one-ai-supercomputer/","published_at":"2026-01-05","accessed_at":"2026-09-02","evidence_type":"vendor-claim","scope":"Official component imagery and vendor descriptions of Vera, Rubin, NVLink 6, ConnectX-9, BlueField-4, and Spectrum-6 roles.","qualification":"The article was updated in March 2026 to add a seventh platform chip; the reused January Figure 4 remains the original six-chip lineup and must be dated as such."}],"claims":[{"id":"agent-is-loop","claim":"A coding-agent turn can contain many model and tool interactions rather than one prompt and one response.","evidence_state":"representative","confidence":"high","source_ids":["openai-agent-loop","anthropic-managed-agents","codex-core"],"qualification":"Implementations differ; the trace sequence is representative, not universal.","proposed_use":"Central thesis and Figure 1."},{"id":"agent-loops-can-scale","claim":"Agent programs can contain tens or hundreds of model calls, depending on the workload and control flow.","evidence_state":"derived","confidence":"high","source_ids":["autellix"],"qualification":"Autellix reports workload-specific distributions, including an average of 159.7 calls for its LATS tree-search programs; this is not a typical count for Codex or coding agents.","proposed_use":"Explain why a user-visible turn can expand into a long systems loop."},{"id":"harness-is-infrastructure","claim":"The harness constructs context, maintains state, routes tools, applies policy, and processes results.","evidence_state":"representative","confidence":"high","source_ids":["openai-harness","openai-safety"],"qualification":"Exact service boundaries and production latency are not public.","proposed_use":"Explain work between model calls."},{"id":"environment-is-physical","claim":"Tool calls become operating-system processes, file operations, runtimes, compilers, tests, and sometimes network services.","evidence_state":"observed","confidence":"high","source_ids":["mts-trace","codex-exec","swe-rex"],"qualification":"The captured fixture measures only search, read, compile, edit, and test work.","proposed_use":"Figures 1 through 3."},{"id":"device-boundary-unknown","claim":"The client-visible model boundary does not reveal provider accelerator type, utilization, queueing, memory traffic, or energy.","evidence_state":"unknown","confidence":"high","source_ids":["openai-agent-loop","mts-trace"],"qualification":"Use model inference at the API boundary; do not label captured time as GPU time.","proposed_use":"Hero disclosure and methodology."},{"id":"cpu-gpu-complement","claim":"Model inference and environment execution depend on different but complementary resources; end-to-end completion emerges from their sequence and handoffs.","evidence_state":"inferred","confidence":"high","source_ids":["openai-agent-loop","openai-harness","mts-trace"],"qualification":"No universal CPU/GPU time share follows from one trace or public documentation.","proposed_use":"Final thesis and critical-path framing."},{"id":"sandbox-varies","claim":"A sandbox is a security boundary, not a synonym for a container; public implementations include OS policies, containers, and virtual machines.","evidence_state":"representative","confidence":"high","source_ids":["openai-safety","anthropic-sandboxing","anthropic-containment"],"qualification":"The MTS fixture used an isolated directory, not a measured container or VM.","proposed_use":"Sandbox explanation."},{"id":"benchmark-gap","claim":"Task-success and inference benchmarks leave much of the agent infrastructure critical path unmeasured.","evidence_state":"inferred","confidence":"high","source_ids":["swe-bench","swe-bench-verified","agentperf","mlperf","agentsysbench"],"qualification":"AgentSysBench is closer to full-loop measurement but is a recent, unreproduced preprint.","proposed_use":"Benchmark matrix."},{"id":"vera-specs","claim":"Vera is an 88-core, 176-thread CPU with up to 1.5 TB LPDDR5X memory, 1.2 TB/s memory bandwidth, 164 MB L3, and 1.8 TB/s coherent NVLink-C2C.","evidence_state":"vendor-reported","confidence":"high","source_ids":["nvidia-vera","nvidia-olympus"],"qualification":"Vendor specifications; not independent agent-workload results.","proposed_use":"Vera architecture box."},{"id":"vera-rubin-role","claim":"In Vera Rubin NVL72, Rubin is the accelerator side and Vera is the host CPU; the vendor describes 72 Rubin GPUs and 36 Vera CPUs.","evidence_state":"vendor-reported","confidence":"high","source_ids":["nvidia-rubin","nvidia-vera"],"qualification":"Preliminary vendor platform specifications.","proposed_use":"Vera/Rubin distinction."},{"id":"vera-sandbox-result","claim":"NVIDIA reports up to 1.8x sandbox throughput for Vera versus an unnamed x86 baseline.","evidence_state":"vendor-reported","confidence":"medium","source_ids":["nvidia-vera-agents"],"qualification":"Configuration-poor vendor benchmark; do not restate as a general agent-speed claim.","proposed_use":"Evidence ladder, with warning."},{"id":"share-unknown","claim":"There is no defensible public universal ratio for agent CPU/GPU time, energy, cost, or critical-path share.","evidence_state":"unknown","confidence":"high","source_ids":["mts-trace","agentsysbench","mlperf"],"qualification":"Those denominators are not interchangeable and require end-to-end telemetry.","proposed_use":"Data-center implications and open questions."}],"benchmarks":[{"name":"SWE-bench","category":"Task success","url":"https://www.swebench.com/","coverage":{"task_success":"yes","model_quality":"partial","inference":"no","tool_calls":"partial","environment":"partial","hardware":"no","critical_path":"no"},"gap":"Pass/fail does not reveal time, utilization, cost, or the component on the critical path."},{"name":"SWE-bench Verified","category":"Human-validated task success","url":"https://openai.com/index/introducing-swe-bench-verified/","coverage":{"task_success":"yes","model_quality":"partial","inference":"no","tool_calls":"partial","environment":"partial","hardware":"no","critical_path":"no"},"gap":"Task validity improves; infrastructure observability does not."},{"name":"SWE-ReX","category":"Execution substrate","url":"https://github.com/SWE-agent/SWE-ReX","coverage":{"task_success":"no","model_quality":"no","inference":"no","tool_calls":"partial","environment":"yes","hardware":"partial","critical_path":"no"},"gap":"Provides reproducible environments but is not itself an end-to-end benchmark."},{"name":"AA-AgentPerf","category":"Inference platform","url":"https://artificialanalysis.ai/methodology/agentperf","coverage":{"task_success":"no","model_quality":"partial","inference":"yes","tool_calls":"partial","environment":"no","hardware":"partial","critical_path":"partial"},"gap":"Uses live inference but simulated tool delays, so the physical environment side is absent."},{"name":"MLPerf Inference","category":"Accelerator inference","url":"https://github.com/mlcommons/inference","coverage":{"task_success":"no","model_quality":"partial","inference":"yes","tool_calls":"no","environment":"no","hardware":"yes","critical_path":"no"},"gap":"Precisely measures bounded inference queries, not a stateful model–environment loop."},{"name":"AgentSysBench","category":"Early system characterization","url":"https://arxiv.org/abs/2608.15127","coverage":{"task_success":"partial","model_quality":"partial","inference":"partial","tool_calls":"partial","environment":"yes","hardware":"partial","critical_path":"yes"},"gap":"Closest public full-loop study found, but recent, unreproduced, mixed local/API visibility, and code-unreleased at review time."}]}