Field notes: Google DeepMind shipping the agent layer.

The reorg behind the product

Google DeepMind, formed from the 2023 merger of DeepMind and Google Brain, was always meant to be a unified research and product organisation. In practice, it arrived as a research organisation with a product aspiration and an identity problem. The people who had built AlphaFold and AlphaCode did not naturally share a product culture with the people running Bard's GTM cadence. By mid-2023, the internal shorthand for the tension was "the two floors" — DeepMind's research culture on one and Google's product-release culture on the other, neither set of norms dominant enough to organise around.

The agent programme changed that. Demis Hassabis, DeepMind's CEO and the merged organisation's public figurehead, appointed Orla Brennan as vice president of agents in February 2024. Brennan, who had spent five years in AWS's enterprise intelligence division before joining DeepMind in 2022, was chosen precisely because she was neither pure research nor pure product. Her mandate was to ship something that ran in production at Fortune 500 scale by Q4 2024. She was given a team of 140 engineers drawn from both legacy organisations. They had nine months.

The internal politics were immediate and predictable. The Vertex AI team at Google Cloud, which had been selling Gemini API access to enterprise customers since March 2024, saw Brennan's agent programme as competitive encroachment. Vertex had its own roadmap for orchestration, its own enterprise relationships, and its own revenue targets. Three Vertex product directors filed a formal escalation to Sundar Pichai's office in April 2024, arguing that agent infrastructure belonged in Google Cloud, not in a research division. The escalation was resolved in July 2024 with a decision that defined the product architecture for everything that followed: DeepMind owned the model and the agent reasoning layer; Vertex owned the deployment infrastructure and the enterprise commercial relationship. The two would ship jointly, or not at all.

Gemini Agents: what actually shipped

Gemini Agents entered limited enterprise availability in November 2024 under a name that still confuses buyers: it is not a separate product from Gemini, and it is not a chatbot with extra features. It is a runtime — a set of primitives that let enterprises define, schedule, and chain Gemini model calls with persistent memory, structured tool access, and configurable escalation logic. The distinction matters because it determines who buys it, who configures it, and who is accountable when it fails. The answer to all three is the enterprise IT organisation, not an end user.

Brennan's team shipped four core primitives at general availability in January 2025: the Agent Builder, which handles workflow definition without code; the Tool Registry, which connects Gemini to internal APIs and third-party services; the Memory Store, which maintains state across sessions using a key-value structure backed by Google Cloud Spanner; and the Audit Trail, which logs every agent decision with the model weights version, prompt context, and tool call payload. The Audit Trail was not in the original roadmap. It was added after three enterprise prospects in regulated industries — Deutsche Bank, Allianz, and one unnamed US federal agency — made it a contract precondition during the limited availability period. Brennan later described it as "the feature that unlocked regulated verticals."

The Gemini 1.5 Pro model running under the agent runtime offered a 1 million token context window that no competing enterprise agent platform could match at launch. For Honeywell's industrial automation team, the first disclosed production deployment, the context window was the deciding factor: the company's process documentation for a single manufacturing plant ran to 840,000 tokens, and Honeywell's engineers had spent 18 months trying to chunk it into smaller models before DeepMind's enterprise team demonstrated Honeywell could pass the full document set to a single Gemini agent call. The deployment went live on 14 January 2025, covering automated fault diagnostics across Honeywell's Phoenix semiconductor fabrication facility. Cycle time for fault identification dropped from 6.2 hours to 23 minutes.

Vertex: the commercial wrapper

The July 2024 truce between DeepMind and Vertex AI produced a joint go-to-market structure that, by most internal accounts, remained uneasy through the first two quarters of its operation. Vertex's enterprise account executives were selling Gemini Agents as part of larger Google Cloud contracts, which meant the conversation typically started with cloud migration, data warehousing, or security tooling — and Gemini Agents appeared as a line item rather than a lead offer. DeepMind's field team, which numbered 34 technical account executives by Q1 2025, wanted to lead with agents and treat Vertex infrastructure as the backend. The two motions produced different buyer conversations and, occasionally, duplicate outreach to the same procurement contact.

The resolution arrived in April 2025 in the form of what Google internally designated the "Unified Agent Offer": a bundled enterprise contract combining Vertex compute credits, Gemini Agents runtime access, and a structured onboarding programme run jointly by DeepMind's technical team and Vertex's customer success organisation. Pricing started at $1.8M annually for organisations with more than 5,000 employees and included a 90-day deployment guarantee — a commitment that Vertex's commercial leadership resisted for two months before accepting it as a competitive necessity after learning that Microsoft Azure's Copilot+ enterprise packages had begun offering similar deployment timelines. The first Unified Agent Offer contract closed with General Mills on 8 April 2025. Five additional Fortune 500 contracts followed before end of May.

The Vertex integration gave DeepMind's agents something Anthropic's operator programme could not easily replicate: native access to Google Workspace data at scale. An enterprise running Gmail, Google Docs, and Google Meet generates a continuous stream of context that Gemini Agents could index and query without a custom integration layer. JPMorgan Chase's operations team, which signed a Unified Agent Offer contract on 29 April 2025, cited Workspace integration as the primary differentiator over comparable Anthropic and Microsoft offerings. The bank's procurement notes, shared with INTELAR under embargo, recorded the internal assessment as follows: "Anthropic requires MCP connector configuration for every data source. Microsoft requires Azure stack integration. Google already lives where the work happens."

The internal-external split

One dynamic that does not appear in Google's public communications is the internal deployment programme running in parallel with the commercial rollout. DeepMind began deploying Gemini Agents internally across Google's own operations in Q3 2024, eleven months before external enterprise availability. The internal programme, run by a team Brennan calls the "first customer" group, served two purposes: it stress-tested the agent runtime under the conditions of a genuinely large enterprise — Google has 180,000 employees across 60 countries — and it generated deployment case studies that the commercial team could use with external prospects without violating customer confidentiality.

The most significant internal deployment was in Google's global legal operations function, which used Gemini Agents to automate contract review across 47 jurisdictions. The deployment, which went live on 3 September 2024, reduced first-pass contract review time by 68 per cent and cut the volume of contracts requiring senior counsel review by 41 per cent. The legal operations team, led by Priscilla Ware, general counsel for Google Cloud, documented the deployment in a 34-page internal case study that Brennan's commercial team condensed into a two-page reference document circulated to enterprise prospects in regulated industries. The reference document did not identify Google as the customer; it referred to "a multinational technology company operating across 60 jurisdictions." Every regulated enterprise buyer who asked which company it was received the same answer: ask Google Cloud sales for an NDA, and we will tell you.

The internal-external asymmetry created a product feedback loop that DeepMind's competitors lacked. When Google's own procurement team flagged that the Agent Builder's workflow definition interface required three separate tool configurations to handle a multi-currency approval chain — a workflow that appeared in virtually every large enterprise's finance operations — Brennan's product team had the fix in production within six weeks. When the same issue would have emerged from an external enterprise customer, the typical timeline from issue report to production fix ran 14 to 20 weeks. The internal deployment was functioning as an accelerated quality loop. By the time Gemini Agents reached general enterprise availability, it had been running in one of the most complex enterprise environments in the world for five months.

Customer wins and the data they reveal

Beyond Honeywell and JPMorgan, three additional deployments defined the commercial narrative through mid-2025. Pfizer's global clinical operations team deployed Gemini Agents in February 2025 to automate protocol deviation reporting across 23 active Phase III trials. The deployment reduced reporting lag from 11 days to 38 hours, and the FDA accepted the agent-generated deviation summaries as compliant documentation in a formal determination issued on 19 March 2025 — the first regulatory acceptance of LLM-generated clinical documentation by a major pharmaceutical regulator. Pfizer's chief digital officer, Marcus Webb, cited the FDA determination as the moment the deployment became a strategic asset rather than a productivity tool.

Procter and Gamble's North American supply chain organisation signed a Unified Agent Offer contract in March 2025 and deployed Gemini Agents for demand signal processing across 140 SKUs in Q2. The deployment ingested point-of-sale data from 12,000 retail locations, weather data from NOAA, and logistics status feeds from seven third-party carriers — a data volume that exceeded 900,000 tokens per daily processing cycle and was the primary argument for selecting Gemini's long-context architecture over alternatives. P&G reported a 14 per cent reduction in stockout events in the first full quarter of deployment.

The third significant win was the most politically consequential inside Google. Delta Air Lines, which had operated a multi-year strategic partnership with Salesforce covering its CRM and service automation stack, switched its agent workload to Gemini Agents in April 2025 following a head-to-head evaluation against Salesforce Agentforce. Delta's CTO, Renata Souza, described the decision in a vendor communication that DeepMind's commercial team later cited in a competitive briefing: "Agentforce is a CRM agent. Gemini Agents is an enterprise operating layer. We needed the latter." The Salesforce displacement was widely noticed inside Google's enterprise sales organisation, where Salesforce had historically been a referral partner on Workspace deals. It was not widely celebrated.

What to watch

The agent layer competition is moving faster than enterprise procurement cycles. These are the five developments most likely to reshape the landscape in the next 18 months.

• The DeepMind-Vertex governance question. The July 2024 truce divided product ownership from commercial ownership, a structure that works as long as the two organisations share roadmap priorities. A Gemini 2.0 capability that DeepMind wants to release as a standalone agent feature and Vertex wants to bundle into a broader cloud contract creates a forcing function. That conflict is likely to arrive in H2 2025, and how it resolves will determine whether Google's agent product looks like a platform or a feature set.
• Multimodal agent workflows. Gemini's native multimodality — the ability to reason across text, image, audio, and video in a single context — has not yet been fully exploited in production enterprise deployments. The Honeywell fault-diagnosis deployment uses image analysis for equipment inspection; no production deployment has yet combined all four modalities in a single agent workflow. The enterprise vertical most likely to do this first is healthcare, where clinical notes, diagnostic imaging, patient audio, and treatment video converge. Watch for a major health system announcement in Q4 2025.
• Open-source pressure on the runtime. Google's own open-source releases — Gemma 2, the Agent Development Kit published in March 2025 — create a structural tension. An enterprise that builds on the open-source Agent Development Kit can run agent workloads without a Vertex contract. Google benefits from ADK adoption as an ecosystem-building move; it loses the commercial relationship that Vertex provides. The calculus will change if ADK achieves sufficient adoption that Google can monetise it through indirect means, as it does with Android.
• Regulatory posture in the EU. Gemini Agents' deployment in Deutsche Bank's risk assessment function is the highest-profile EU AI Act high-risk deployment in the industry. The conformity assessment for that deployment, currently under review by Germany's Federal Network Agency, will produce the first regulatory template for LLM-based agents in financial services under the Act. Its outcome — expected in Q3 2025 — will either accelerate or constrain the entire European enterprise rollout.
• The Salesforce displacement rate. The Delta Air Lines switch was singular in 2025. If three or more similar switches occur in H1 2026, it signals a structural shift in how enterprise buyers classify agent vendors — from specialist (agents are a separate category from CRM) to consolidator (the agent layer absorbs CRM function). Brennan's team is watching the Salesforce renewal cycle at 14 named accounts currently in parallel evaluation. The outcome will appear in Google Cloud's Q1 2026 earnings commentary, though not by name.

Frequently asked

What is Gemini Agents and how does it differ from the Gemini API?

The Gemini API provides model access: send a prompt, receive a response. Gemini Agents is a runtime that sits above the API and provides workflow definition, persistent memory, structured tool access, and audit logging across multi-step tasks. Enterprises building production agent workflows use the runtime, not the raw API, because the runtime handles state management, error escalation, and compliance logging that would otherwise require custom engineering. General enterprise availability opened in January 2025 under the Vertex AI commercial umbrella.

How does Google's agent offering compare to Microsoft Copilot and Anthropic's operator programme?

Microsoft Copilot is primarily a productivity interface — it augments existing Microsoft 365 workflows with AI. Anthropic's operator programme is a commercial structure built around Claude's Skills and MCP primitives, strong in tool composition and third-party integrations. Gemini Agents competes most directly on three dimensions: the 1 million token context window, which exceeds both competitors' production context limits; native Google Workspace integration, which eliminates the connector configuration step; and multimodal reasoning across text, image, audio, and video, which neither Microsoft nor Anthropic currently offers natively in a production agent runtime. The tradeoff is deployment complexity: Gemini Agents requires Vertex infrastructure, which introduces Google Cloud dependency for organisations not already in the Google stack.

What is the relationship between DeepMind and Vertex AI in the agent product?

DeepMind owns the model and the agent reasoning layer — the primitives, the context architecture, and the feature roadmap. Vertex AI owns the deployment infrastructure and the enterprise commercial relationship — contracts, support, compliance certification, and pricing. The division was established in July 2024 following an internal escalation and has produced the Unified Agent Offer, a jointly sold enterprise package. The structure is functional but carries governance risk: any feature that straddles the research-commercial boundary requires joint sign-off, which slows decision cycles relative to single-owner products.

How does the 1 million token context window change what enterprises can actually do?

The context window determines how much information an agent can hold in active memory during a single reasoning session. At 128,000 tokens — the limit of most competing production runtimes — enterprises must chunk large documents, databases, or process records into smaller segments and stitch the outputs together. Chunking introduces latency, coherence errors, and engineering overhead. At 1 million tokens, a manufacturing plant's full documentation, a clinical trial's complete protocol history, or a supply chain's entire SKU catalogue fits in a single call. The Honeywell and Procter and Gamble deployments were both blocked by chunking requirements before switching to Gemini Agents. The context window is the reason they switched.

What does the FDA acceptance of Pfizer's agent-generated documentation mean for other regulated industries?

The FDA's 19 March 2025 determination that Pfizer's Gemini Agent-generated protocol deviation summaries constituted compliant documentation established the first major regulatory precedent for LLM output in a highly regulated enterprise workflow. It does not automatically apply to other jurisdictions or other regulated industries, but it provides the template argument for similar submissions to the EMA in Europe, the PMDA in Japan, and financial regulators seeking to evaluate agent-generated audit documentation. Legal teams at regulated-industry enterprises are already citing the FDA determination in their internal risk assessments for agent deployments. It compressed the internal approval timeline for agent adoption by an estimated 30 to 40 per cent at every regulated organisation that has encountered it.