Google

Confidence 0.90 · 11 sources · last confirmed 2026-06-18

Big-tech platform company; operating subsidiary of Alphabet. Within this wiki Google appears primarily in three roles: AI lab (Gemini family of foundation models, ADK harness framework), cloud hyperscaler (Google Cloud Platform and the Agent Platform that hosts agent-runtime infrastructure), and research publisher (Google Research preprints and reports). The wiki holds a small but growing primary-source footprint on Google as of May 2026 — see Mentioned in below for current edge count.

Sub-organizations and brands appearing in the wiki

• Google Research — research arm; producer of the Gemini model family; anchor of the durable-skills measurement work (Globerson et al. 2026 / Vantage / Executive LLM platform).
• Google Cloud / Google Cloud Platform — hyperscaler division. Surfaces in this wiki via Cloud Run, GKE, Agent Runtime, Cloud Build, Cloud Trace / Logging / Monitoring as deployment, CI/CD, and observability substrates for agents. Sponsor of the Agents CLI announcement. Also the publisher of the Open Knowledge Format (OKF) v0.1 (McVeety & Hormati 2026) — an open spec formalizing the LLM Wiki pattern, with Google Cloud’s Knowledge Catalog updated to ingest it. OKF is a second instance of the Bfloat16 pattern noted below — a Google-built data/knowledge format proposed as an open industry standard.
• Google Developers — developer-relations division publishing the Google Developers Blog; the channel for the Agents CLI announcement.
• Gemini Enterprise — enterprise distribution surface for agents registered via the Agents CLI. (Single-source mention; promote on second-source coverage.)
• Google Brain — Google’s ML research division (subsumed into Google DeepMind / Google Research over time). Surfaces in this wiki as the originator of Bfloat16 — the ML-bespoke 16-bit floating-point format (1 sign + 8 exponent + 7 mantissa) that won the training-precision war over IEEE 754 FP16 (1+5+10) because the extra exponent bits give wider dynamic range at the cost of mantissa precision. Bfloat16 is named in Turc 2025 as “introduced by Google Brain, hence the B, specifically for machine learning … important for training models where we want the ability to represent very low numbers in the vicinity of zero.” The wiki’s first vendor-built-data-format-becomes-industry-standard example.

Edge ML hardware

• Coral Edge TPU — Google’s low-power processor for IoT devices (sensors, smart cameras, on-device ML). Named in Turc 2025 as the example of edge-ML hardware that “doesn’t even support floating-point operations at all” — making INT8 quantization not an optimisation but a prerequisite for deployment. The wiki’s first Google edge-ML hardware mention, and the wiki’s first concrete anchor for the bottom-end of the FTSG edge-AI convergence claim.

ML frameworks

• TensorFlow — Google’s open-source ML framework. Named in Turc 2025 as having built-in quantization support alongside Meta’s PyTorch; both abstract the integer-only-arithmetic / fixed-point-multiplication machinery from end-user model developers. The framework-level companion to the Coral hardware-level edge-ML deployment story.

Platform / product engineering

• Agent Platform — Google Cloud’s umbrella for agent-runtime infrastructure.
  • Agents CLI in Agent Platform — released 22 April 2026 as the unified programmatic backbone for the Agent Development Lifecycle (ADLC); bundles seven specialized skills (Workflow / ADK Code / Scaffold / Evaluation / Deployment / Publish / Observability); two-mode design (Agent Mode for AI coding agents, Human Mode for deterministic CLI use). See 2026-04-22-cheung-ippolito-secchi-google-agents-cli.
  • Agent Runtime / Cloud Run / GKE — three deployment substrates the CLI targets via agents-cli deploy.
• ADK (Agent Development Kit) — Google’s open-source agent-engineering framework. Listed in Kokane’s Layer 2 harness-frameworks alongside LangChain, Microsoft Agent Framework, LlamaIndex, CrewAI, Haystack, DSPy. The Agents CLI’s ADK Code skill is the Python-API reference surface for ADK. The Google Cloud Tech first-agent tutorial (June 2026) is the wiki’s first hands-on ADK build: a self-correcting multi-agent blog-writer using LlmAgent / LoopAgent / root-agent-with-tools, framed explicitly on the ReAct loop.
• A2A / A2UI / MCP integration — Google Cloud names A2A and A2UI alongside MCP as the agent-integration protocols the platform supports natively.

Models referenced in this wiki

• Gemini family — the Google Research foundation-model line referenced as the autorater substrate in Globerson et al. 2026 (Pearson r = 0.88 vs human experts on creativity assessment).
• Gemini CLI — Google’s coding-agent CLI; named as an Agents-CLI consumer alongside Claude Code and Cursor in the Agents CLI announcement.
• Gemini API File Search — hosted RAG primitive on the Gemini API. May 2026 update added multimodal support (via Gemini Embedding 2), custom metadata filters, and page-level citations. See 2026-05-05-google-gemini-file-search-multimodal.
• Gemini Embedding 2 — Google’s multimodal embedding model (text + images), powering File Search as of May 2026.
• gemini-3-flash-preview — generation model demonstrated in the File Search code sample.
• NotebookLM (notebooklm.google) — Google’s note-taking / multi-source-research tool with audio-overview and (newer) video-overview features. Surfaces in Nika 2025 as an unconventional AI-as-judge pattern: upload all student-pitch audio recordings into one notebook, prompt the audio-overview hosts with judging criteria, generate a live interactive radio-podcast that announces winners. Two-way interactive mode allows audience call-ins.
• Flow (Google Labs, labs.google/flow) — text-to-video creator app, powered by Veo. Used in Nika 2025 for the smart-fridge promotional clip generation.
• Veo — Google’s video-generation model, the engine behind Flow. Also referenced in passing as a tool a PM can prompt directly with PRD-derived content.
• Google AI Studio (prototyping) — named in Nika 2025 alongside v0 and Lovable as a vibe-coding-class prototyping tool. Not deeply demonstrated.

People at Google referenced in this wiki

• Marily Nika — AI Product Lead at Google; founder of the AI Product Academy (separate from her Google role). First wiki source mention by name in Nika 2025 (How I AI). Per the author-entity-promotion rule, do not promote on a single source.
• James Manyika — see James Manyika entity (already in the wiki).
• Jaana Dogan — engineering leader on Gemini API, surfaced via Kiron-Schrage 2026’s anchor anecdote (“It’s not perfect and I’m iterating on it.”). Single-source first mention; not yet promoted.
• Ivan Solovyev / Kriti Dwivedi — Gemini File Search team members; see Solovyev & Dwivedi 2026.
• Nicole Forsgren — lead, Developer Intelligence team at Google; founder of the DORA (DevOps Research and Assessment) programme; lead author of Accelerate (2018). Speaker at Google I/O 2026 in [[2026-04-21-forsgren-macvean-build-core-skills-thrive-ai-era-developer|Build core skills to thrive as an AI-era developer]]. The talk explicitly cites the DORA research for both the productivity-paradox claim (individual gains, team-level losses — the engineering-team correlate of micro-productivity-trap) and the AI-as-amplifier-and-mirror framing. First-mention; promote on second-source coverage.
• Andrew Macvean — lead, Developer Intelligence team at Google; co-presenter with Forsgren at Google I/O 2026. First-mention; promote on second-source coverage.

Internal teams referenced in this wiki

• Developer Intelligence team — Google’s in-house engineering-research team studying how Google engineers work with AI. Anchor for Forsgren & Macvean 2026. The talk’s data points — “~3/4 of all code at Google is now written by AI” and “no measurable increase in outages or decrease in system reliability” — are this team’s research output.
• DORA (DevOps Research and Assessment) — Forsgren-led research programme running annual State of DevOps reports since 2013; the empirical backbone for Forsgren & Macvean 2026’s productivity-paradox and amplifier-and-mirror claims. First-mention as a wiki referent; promote to its own entity page on second-source coverage.

Inside-Google engineering practices referenced in this wiki

The Macvean 2026 talk anchors several concrete Google practices worth tracking as discrete referents:

• Search-org PM platforms. Internal platforms that let PMs ship features all the way through to live production experiments without explicitly authoring each line of code — the Google-side empirical exemplar for Singhal’s PM-as-product-builder thesis.
• Code-review agent stack. Three-layer agent system used at fleet scale: Code Review Agents (functionality / reliability / performance / usability / security / maintainability), Shepherding Agents (guide changes through CI/CD; trigger reviewers multiple times), Risk Assessor Agents (scan in-flight changes; flag high-risk for human review).
• TensorFlow migration three-agent architecture. A planner / orchestrator / coder triad with product-specific playbooks (YouTube-specific style for YouTube-area migrations). Cited as Google’s canonical multi-agent-system exemplar.
• Quality Agent. Stress-tests requirements and looks for logical collisions before execution begins. Single-mention as of this ingest; promote on second-source coverage.

Google DeepMind

The combined Google AI research org (formed via DeepMind / Google Brain merger). Surfaces here as the publishing org for the Gemini File Search multimodal announcement (Solovyev’s team affiliation). Distinct from Google Research (the older Research arm), though boundaries are not always clear in public-facing communications. Promote to its own entity page on second-source mention if the editorial distinction becomes load-bearing.

Open questions

• Cross-product coherence. Two Google primaries in the wiki within ten days (Globerson et al. 2026 from Google Research; Cheung et al. 2026 from Google Developers) hint at a broader Google agent-platform thesis. Whether Google’s positioning is Gemini-as-substrate + agent-platform-as-runtime + assessment-as-product or something looser will become clearer with more sources.
• ADK as concept page. ADK is referenced from three sources but currently treated as a brand mention on this entity. If a primary ADK technical-deep-dive lands, ADK could justify its own concept (or product-entity) page.
• Gemini family entity. The wiki has Anthropic’s Claude family enumerated explicitly (Sonnet 4.5 / Opus 4.5 / Opus 4.6) but does not yet have the parallel for Gemini. Worth filling on second-source coverage.