In this article we will analyze how Google, OpenAI, and Anthropic are productizing ‘agentic’ capabilities across computer-use control, tool/function calling, orchestration, governance, and enterprise packaging.
Now, agent platforms are the key to competitive advantage, and not just models. Google is aligning Gemini 2.0 with an enterprise control plane on Vertex AI and a new ‘front door’ called Gemini Enterprise. OpenAI is consolidating developers early around Responses API. They are packaging lifecycle agent elements into AgentKit. And they’re deploying an all-purpose GUI controller, the Computer-Using Agent. Anthropic expands Computer Use, while Artifacts is transformed into a lightweight tool for internal rapid tools.
OpenAI: CUA, AgentKit, and Responses for Agent Surface.
Computer-using Agent (CUA).
OpenAI launched Operator powered by CUA in January 2025. CUA uses GPT-4o class vision combined with reinforcement learning to create GUI policies. It executes using human-like development such as mouse and keyboard. The stated purpose is a single interface that generalizes across web and desktop tasks.
API Responses
OpenAI redefined Responses to be the agent-native API. The new design integrates chat, multimodality, tool use and state into one initial step. This simplifies the historical split across Chat Completions and Assistants, formalizing hosted tools and persistent reasoning in a single endpoint.
AgentKit
Launched in October 2025, AgentKit packages agent building blocks: visual design surfaces, connectors/registries, evaluation hooks, and embeddable agent UIs. This will reduce the complexity of orchestration and standardize all aspects of agent deployment, from design to implementation.
Risk Profile
Third-party testing has revealed brittleness in practical automations. These include flaky DOM target, loss of window focus, and failure to recover after layout changes. OpenAI isn’t the only company that has this problem, but SLAs for production are affected. Teams must implement high-risk tasks behind reviews, gate them, and stabilize selectors. Pair CUA experiments with execution-based evaluation such as OSWorld tasks.
PositionOpenAI optimizes a programmable agents substrate: an API surface (Responses), lifecycle kit(AgentKit), universal GUI controller(CUA). For teams willing to own their evaluation harness and operations, this stack provides tight control and fast iteration loops.
Gemini Enterprise and Astra 2.0 for Perception. Vertex AI agent builder for orchestration.
Models Runtime
Google frames Gemini 2.0 as ‘built for the agentic era,’ with native tool use and multimodal I/O including image/audio output. Project Astra demos show low-latency perception, continuous assistance and patterns of always-on perceptual feedback that are mapped to loops for planning as well as acting. These capabilities are intended to feed Gemini Live and the broader agent runtime.
Vertex AI Agent Builder
Vertex AI agent builder is Google’s control plan for creating and deploying GCP agents. The Google Vertex AI agent builder is the control plane that allows you to create and deploy agents using GCP. official documentation Vertex integration is demonstrated by the Agent Garden, which includes tools and templates for creating multi-agent interactions, as well as orchestration. This serves as the platform to implement policies, logging, and evaluation pipelines for GCP users.
Gemini Enterprise
In October 2025, Google announced Gemini Enterprise as a governed front door to ‘discover, create, share, and run AI agents’ with central policy and visibility. The emphasis is on cross-suite integration spanning Google Workspace, Microsoft 365/SharePoint and line-of business integrations like Salesforce and SAP. This is positioned as a fleet-level governance layer, not only a development kit.
Application Surface
Google has also pushed agentic controls into the end-user environment. Agent Mode, in both the Gemini App and Project Mariner, extends consumer and professional workflows. It includes teach-and-repeat functionality, task management and auto-execution for tasks as common as searching and filtering. This serves as both a data source for guardrails and a proving ground for UI-safety patterns.
PositionGoogle optimizes for enterprise-wide deployments with surface integration. If you need centralized policy/visibility across many agents, with Workspace and cross-suite context, the Gemini Enterprise + Vertex pairing offers the most prescriptive path today.
The Anthropic View of Computer Use, App Builder Path and Artifacts
Computer Use
Anthropic launched Computer Use for Claude 3.5 Sonnet explicitly in October 2024 as a Beta capability, which requires the appropriate software set-up to mimic human cursors and keyboard interactions. The company was very open about the error profiles that it created and its need for careful mediating. For production, expect policy-first defaults and incremental broadening rather than a hard pivot to full autonomy.
Artifacts → App Building
Anthropic will extend Artifacts to Claude in June 2025. This feature allows users to create, share, and host interactive apps from Claude. This feature is aimed at rapid internal apps and mini-apps that can be shared. Developers can create apps that call back into Claude via a new API, and published app usage bills the end user rather than the author.
PositionAnthropic optimizes its creation for human-inthe-loop with an explicit safety posture. The combination of Computer Use and Artifacts supports a design pattern where users co-pilot agents, validate actions, and graduate prototypes into shareable internal apps without heavy scaffolding.
Selecting Agents: The Benchmarks that Matter
Use the Function/Tool Number
Berkeley Function-Calling Leaderboard V4 goes beyond just single calls. It includes multi-turn scheduling, live/nonlive settings and hallucination measurements. You can use BFCL for tool-routing quality, argument fidelity, and sequencing under state changes.
Web/Computer Use
OSWorld has defined a benchmark consisting of 369 desktop tasks, based on execution-based evaluations for multiple OSes. Original results showed large human–agent gaps and identified GUI grounding as a major bottleneck. You can treat OSWorld as the minimum bar for assessing GUI agents, then layer domain-specific workflows.
Conversational Agents
τ-Bench simulates dynamic conversations where an agent must follow domain rules and interact with tools; the 2025 τ²-Bench extension adds dual-control scenarios where both the user and agent can act, increasing realism for support workflows. You can use these when you care about policy adherence, user guidance, and multi-trial reliability.
Software Engineering Agents
Leaderboards for the SWE Bench family cover issue resolution from beginning to end. SWE Bench Pro (2025), with its 1,865 instances spread across 41 repositories, increases task difficulty while adding contamination resistance. For engineering assistants, you should not rely on ‘Lite’ alone—run Verified or Pro with a locked scaffold.
Comparison Analysis
Model Core and Modality
OpenAI’s current implementation combines GPT-5 orchestration using Responses, with a CUA (general GUI controller). The result is a unified interface for both reasoning and tool use, as well as a RL-trained controller for actions on screen. Google will push Gemini 2.0 for multimodal low-latency perception, and Astra to expose agent plumbing. Computer Use is a new feature from Anthropic, which enables Claude 3.5. Artifacts allows users to turn prompts into apps and call models. The differences map to strategy: programmable substrate (OpenAI), governed enterprise scale (Google), and human-in-the-loop app creation (Anthropic).
Agent Platform and Lifecycle
OpenAI AgentKit, a toolkit with an informed opinion, reduces the need for custom scaffolds while aligning itself to Responses. Google’s Vertex Agent Builder provides multi-agent coordination and governance hooks within a GCP native control plane. Anthropic’s Artifacts/app-builder anchors a rapid prototyping loop for internal tools and user-validated workflows. Select based on where you want to spend engineering effort: programmable pipelines (OpenAI), centralized IT management (Google), or fastest human-supervised iteration (Anthropic).
Governance and policy
Google Gemini Enterprise provides the most comprehensive fleet governance solution: central policies, visibility, context across Workspace, Microsoft 365 and line-of business apps, as well as connectors. OpenAI consolidation in Responses should reduce integration surfaces, and simplify policy attachment. However, enterprise posture can vary by architecture. Anthropic’s default stance is cautious feature rollout with explicit policy framing and human mediation.
Assessment Story and External Signs
OpenAI claims strong computer-/browser-use performance for CUA, but independent harnesses like OSWorld still report significant gaps across agents. Google’s messaging focuses on demos and enterprise deployments. Vertex can verify the claims made by Google about BFCL and OSWorld workloads. Anthropic’s Artifacts provides a pathway to test-and-deploy small apps quickly, then measure them against τ-Bench-style dialogue tasks and OSWorld-style GUI tasks.
Guide for Deployment of Technical Teams
Lock the Runner Before Model
Use state-aware, execution-based harnesses. Use OSWorld’s task scripts and verified setups for GUI control. For orchestration of tools, use BFCL V4 components such as multi-turns and hallucinations. For policy-bound dialogues, prefer τ/τ²-Bench. SWE Bench Verified or PRO is a good choice for engineers assistants. Keep the runner constant while iterating on models, prompts, and retries.
Choose the Governance Location
Vertex AI AgentBuilder and Google Gemini Enterprise provide the best governance plan if you require centralized visibility over many agents, plus context from Workspace or Microsoft 365. OpenAI’s AgentKit and Responses stack are a good choice if you’re looking for a programable substrate. Anthropic’s approach favors human-in-the-loop controls with clear policy boundaries through the product surface.
The GUI Design Failure and Recovery
Visual similarity, selector drifting, changing window focus, and confusion of detectors can confuse them. You can build retries, add ‘are we on the right page’ checks, and gate irreversible actions behind review. This guidance applies to OpenAI CUA and Anthropic Computer Use alike, and the gaps are documented in OSWorld results.
The Iteration Method is the Best Way to Optimize Your Code
If you prototype many small internal tools, Anthropic’s Artifacts/app-builder minimizes scaffolding and lets non-specialists contribute. AgentKit and Responses offer the best primitives for a deeply programmable, hosted pipeline with tools and memory. For governed, fleet-level rollouts, Google’s Vertex + Gemini Enterprise stack is designed for IT-managed scale.
Vendor Bottom Line
OpenAIThis stack is attractive when you want direct control over tools, memory, and evaluation, as well as being prepared to operate your own runners. This stack can be attractive if you are looking for direct control of your tools, evaluations, and memory. You will also need to have the capability to manage your own runners. You can validate GUI tasks on OSWorld and dialogue planning on τ-Bench.
GoogleVertex AI Agent Builder orchestrator and Gemini Enterprise to provide organization-wide policies, visibility and context across the suite. It may be the simplest way to standardize agent operations for large estates that use Workspace and hybrid 365 environments. You can test tool quality on BFCL and GUI reliability on OSWorld before scaling.
Anthropic: A human-in-the-loop path: Computer Use plus Artifacts/app-builder for rapid creation and sharing of internal apps. It is a good option for teams who want to iterate quickly with clear checkpoints. You can use τ-Bench to assess policy adherence and user guidance, and OSWorld to check GUI action reliability.
Commentary on the Editorial
Three fundamentally distinct philosophies will define enterprise AI in 2025. OpenAI’s decision to use a programmable, unified substrate is in line with their developer-first philosophy, but could be overwhelming for teams that lack strong engineering abilities. Google’s play on enterprise governance is strategic given its Workspace dominance but feels bureaucratic in comparison to nimble, iterative cycles which define successful AI implementations. Anthropic’s human-in-the-loop approach appears most aligned with current organizational realities—where trust, not just capability, remains the bottleneck for AI adoption. Technical superiority may not determine the winner, but rather which vendor is best able to bridge AI possibilities and enterprise realities. You can get started with 95% of generative AI According to MIT Research, if pilots fail to make it to production, then the platform which solves the deployment friction and not just the model performance is likely to capture the biggest share of projected revenues. $47.1 billion AI agent Market by 2030
References:
- https://www.fanktank.ch/en/blog/choosing-ai-models-openai-anthropic-google-2025
- https://www.mindset.ai/blogs/in-the-loop-ep15-the-three-battles-to-own-all-ai
- https://deeplp.com/f/xxx
- https://akka.io/blog/agentic-ai-tools
- https://www.alvarezandmarsal.com/thought-leadership/demystifying-ai-agents-in-2025-separating-hype-from-reality-and-navigating-market-outlook
- https://www.datacamp.com/blog/best-ai-agents
- https://mashable.com/article/best-ai-agents-work
- https://claude.ai/public/artifacts/e7c1cf72-338c-4b70-bab2-fff4bf0ac553
- https://techcrunch.com/2025/01/23/openai-launches-operator-an-ai-agent-that-performs-tasks-autonomously/
- https://openai.com/index/introducing-agentkit/
- https://cloud.google.com/blog/products/ai-machine-learning/introducing-gemini-enterprise
- https://www.anthropic.com/news/3-5-models-and-computer-use
- https://openai.com/index/introducing-operator/
- https://openai.com/index/computer-using-agent/
- https://openai.com/index/new-tools-and-features-in-the-responses-api/
- https://developers.openai.com/blog/responses-api/
- https://techcrunch.com/2025/10/06/openai-launches-agentkit-to-help-developers-build-and-ship-ai-agents/
- https://felloai.com/2025/10/openai-launches-agentkit-for-building-ai-agents-here-is-all-you-need-to-know/
- https://www.technologyreview.com/2025/01/23/1110484/openai-launches-operator-an-agent-that-can-use-a-computer-for-you/
- https://shellypalmer.com/2024/12/google-launches-gemini-2-0-ushering-in-the-agentic-era/
- https://blog.google/products/gemini/google-gemini-ai-collection-2024/
- https://blog.google/technology/google-deepmind/google-gemini-ai-update-december-2024/
- https://techcrunch.com/2025/10/09/google-ramps-up-its-ai-in-the-workplace-ambitions-with-gemini-enterprise/
- https://www.reuters.com/business/google-launches-gemini-enterprise-ai-platform-business-clients-2025-10-09/
- https://blog.google/products/google-cloud/gemini-enterprise-sundar-pichai/
- https://www.anthropic.com/news/developing-computer-use
- https://www.nist.gov/news-events/news/2024/11/pre-deployment-evaluation-anthropics-upgraded-claude-35-sonnet
- https://www.infoq.com/news/2025/06/anthropic-artifacts-app/
- https://www.anthropic.com/news/build-artifacts
- https://www.anthropic.com/news/claude-powered-artifacts
- https://gorilla.cs.berkeley.edu/leaderboard.html
- https://gorilla.cs.berkeley.edu/blogs/15_bfcl_v4_web_search.html
- https://openreview.net/forum?id=2GmDdhBdDk
- https://mlq.ai/media/quarterly_decks/v0.1_State_of_AI_in_Business_2025_Report.pdf
