OpenAI Rebuilds Git. Agent Security Gets a Kill Switch.
Five developments for teams running AI: faster source control, verifiable inference, runtime controls, AMD demand and a chemistry workbench.
This article was researched and written with AI assistance.
AI-generated illustration. Conceptual illustration of the infrastructure and controls around AI work.
November 4, 2026. Coverage: announcements and disclosures from November 2 and 3.
Adding more coding agents can leave everyone waiting for the same repository. OpenAI’s account of its Pando Git backend explains how it handles concurrent writes and repeated downloads, and which consistency guarantees it gives up. Platform teams can learn from the design well before they reach OpenAI’s scale.
The other four stories concern different decisions: checking which model served a response, stopping an agent, evaluating another compute supplier and connecting a scientific assistant to established tools. The reporting draws on company announcements and technical material. Performance figures remain their measurements; the suggested trials are editorial recommendations.
AI-generated illustration. Parallel development puts pressure on repository writes and repeated reads.
OpenAI explains why it rebuilt the Git backend
On November 2, OpenAI described Pando, the backend now serving its main development repository. Production traffic began in April, so this is a new engineering disclosure about an existing internal system. OpenAI reports a reduction of nearly 80% in median push latency after migration. Its engineering account explains the tradeoff.
Pando coordinates updates to individual branches separately and reuses stored packfiles, the compressed bundles Git transfers. Unrelated writers can proceed in parallel while readers download cached bytes. Reading several branches no longer guarantees a single repository-wide snapshot. Updates to the same branch still require coordination.
For a team running agents on separate fixes, each extra worker can trigger another round of repository downloads and CI jobs. Time those stages separately. An agent that finishes its patch quickly may still wait for a checkout, push or build before anyone can review the result. Record that waiting time before buying more model capacity.
Pando remains an internal system. OpenAI says it is exploring ways to share the technology and links to public Git contributions. Its production numbers describe its own repository; the page’s interactive comparison is explicitly a teaching simulation. Neither establishes the speedup another workload would receive.
OpenAI’s public Git tracker lists work on concurrent pack downloads and parallel pack indexing. These client changes support the packfile reuse discussed in its Pando report. Source: OpenAI Git contributors. Enlarge to inspect the fine labels.
AI-generated illustration. A computational proof can bind an output to specified model weights.
Anthropic makes a narrower promise with FlashProve
Suppose an application pays a routing service for a particular model. A plausible answer cannot establish which model produced it. Anthropic’s November 2 FlashProve release explores a cryptographic check: the provider commits to weights, then supplies evidence that a specified computation agrees with them. The design aims to keep those weights private.
The prototype evaluates an integer version of 4-bit Llama 4 Maverick, with contexts up to 1,024 tokens. In the reported exact-match benchmark, each proof checks one output token under greedy decoding. These proofs are produced after inference. The reported throughput counts prefill tokens, so it does not describe an unrestricted chatbot generating verified answers at that speed.
Anthropic’s technical report puts the compute overhead at about 70 to 95 times its vLLM baseline in GPU time. The comparison also uses different numerical formats. Formal protocol work and a Lean verifier accompany the release, but the fast executable implementation does not carry the same formal guarantee.
Model identity and answer quality need separate acceptance tests. A valid proof of computation would still leave a wrong calculation or an unsuitable tool call to catch elsewhere. An evaluator could eventually use this evidence to reduce uncertainty about silent model substitution. A production buyer should first ask which tokens and decoding settings are covered, how long verification takes, and who audited the implementation.
FlashProve’s repository paper reports proof costs under several configurations. The exact-match rate counts prefill tokens; each proof in that row checks one output token after a 1,024-token window. It is not a verified response-generation rate. Source: Anthropic PBC, Table 3, Apache 2.0. Enlarge to inspect the fine labels.
AI-generated illustration. An enforcement control’s location determines which activity it can interrupt.
CrowdStrike’s kill switch starts at the gateway
CrowdStrike announced AI Kill Switch in Falcon Guardian on November 3. Its technical launch post says administrators can direct connected AI gateways to block model access now. The gateway can be CrowdStrike’s or a supported third-party service. Coordinated containment across endpoints, cloud environments, identities, SaaS and browser sessions is described as coming soon.
Consider an agent that has launched a script and is waiting for it to finish. Cutting off its next model request does not establish that the script has stopped. A gateway intervention and a process termination are different operations; each needs its own check.
CrowdStrike’s published Kill Switch screen selects model providers and connected gateways before activation. The November 3 release blocks model access at the gateway; broader endpoint containment remains planned. Source: CrowdStrike. Enlarge to inspect the fine labels.
AI-assisted diagram. Simplified control map based on CrowdStrike’s description. This illustrates the announced scope, not an independently tested product architecture.
Get AI & Data Engineer’s stories in your inbox
Join Medium for free to get updates from this writer.
For teams evaluating the feature, try a small sandbox exercise. Start a harmless task, activate the control, then inspect subsequent model calls and any local work already underway. Check the audit record and the procedure for restoring access. This reporting does not establish independent containment rates or response times. Base a purchase decision on the controls available in the intended environment, and track planned additions separately.
AI-generated illustration. Compute procurement depends on deployment evidence as well as supplier growth.
AMD’s revenue shows demand; workload tests decide suitability
AMD’s November 3 results report USD 13.377 billion in revenue for the quarter ended September 26, up 45% year over year. Data Center revenue reached approximately USD 8.5 billion, up 97%. AMD attributes that segment’s growth to EPYC processors and the launch of Instinct MI450 GPUs.
The company also says its first Helios racks are running models and workloads at customers including Anthropic, Microsoft and OpenAI. The release names deployments but provides no independent workload comparisons. Data Center includes CPUs as well as accelerators, so its growth rate cannot be read as an AI GPU market-share figure.
AMD’s November 3 results slide puts Data Center revenue at $8.5 billion, up 97% year over year. The segment includes EPYC CPUs and Instinct GPUs, so the chart does not isolate AI accelerator sales. Source: AMD, Q3 2026 results, slide 14. Enlarge to inspect the fine labels.
For a team considering a second compute supplier, the next useful evidence is a representative workload on the software stack it expects to maintain. Keep the model, context lengths and quality target fixed; include the cost of porting custom kernels and diagnosing failures. A favorable token rate can lose its appeal if operational work consumes the saving. Request access to the actual deployment configuration before treating a rack announcement as a procurement option.
AMD forecasts roughly USD 17 billion in fourth-quarter revenue, plus or minus 3%. That guidance reflects management’s expectations; the reported quarter supplies completed-period evidence. Neither figure prices a customer’s workload or settles whether its application will run economically on the hardware.
AI-generated illustration. An assistant coordinates a scientific workflow while specialist software performs calculations.
Google puts established chemistry tools inside an agent workflow
Google DeepMind’s November 3 walkthrough introduces Cheminformatics Skills for Antigravity. The assistant coordinates literature and database work while tools such as RDKit handle molecular calculations, with connections to ChEMBL and PubChem. Google says the skills are available through Antigravity’s Science plugin.
Its example starts from the lung-cancer drugs crizotinib and lorlatinib, curates related compounds and explores how molecular rigidity relates to potency. The workflow proposes further designs and synthesis routes. Google reports completing the demonstration in about an hour; no independent general speedup is established here.
Google’s Antigravity walkthrough places the literature search beside a dashboard for 568 curated compounds. The plots are part of Google’s demonstration; candidate molecules still need experimental validation. Source: Google DeepMind. Enlarge to inspect the fine labels.
For teams building similar workflows, preserve the specialist tool and make its inputs and outputs inspectable. A scientist could start with a small, already characterized compound set, inspect the extracted structures against their sources, and rerun the calculations. That would test the data-handling chain before relying on a new design proposal. Comparing a polished report with an accepted answer alone could miss an extraction error that happens to produce a plausible conclusion.
Google explicitly warns about sparse data and differences between laboratory assays. Proposed molecules still require synthesis and experimental validation. Retain the source records and calculation outputs so a scientist can investigate a surprising result.
AI-generated illustration. A small trial is useful when it resolves a specific uncertainty.
Pick the uncertainty that blocks your next decision
If agent work is piling up, start by measuring the wait between a completed change and a reviewable result. If the concern is control, rehearse interruption in a sandbox and record which activities survive it. Each exercise gives a team an observation it can act on without first adopting an entire new platform.
Write down the expected result before running the trial, including what would make you stop. Keep that note beside the logs. Use the recorded requirement to judge whether the next announcement addresses the problem you measured.
Subscribe to Data & AI Engineer for practical analysis of the evidence behind AI tools and infrastructure.
Subscribe to Data & AI Engineer on Substack.
Sources
- OpenAI, Scaling Git at OpenAI, November 2. Engineering disclosure; Pando production use began in April.
- Anthropic, Progress on provable inference for large language models, November 2; technical report, benchmark setup and implementation limitations. The figure above is from the repository paper, Table 3.
- CrowdStrike, Introducing AI Kill Switch, November 3; enterprise security announcement, November 3.
- AMD, Third-quarter 2026 financial results, November 3, for the quarter ended September 26.
- Google DeepMind, Agentic cheminformatics with Google Antigravity, November 3. Product walkthrough and experimental-validation limits.
