Agent-accessible software infrastructure

Enterprise software is exposing more than data access: it is packaging execution, governance and domain context as an agent runtime. MongoDB introduced an engine for building and governing agents directly on operational data; Atlassian and OpenAI are connecting enterprise knowledge to actions; NVIDIA describes controlled network agents adopted by major telecom operators; and OpenAI is evaluating computer-use agents inside contracting workflows. This extends the existing agent-accessible-software line from tool interfaces into managed execution inside systems of record. Most evidence is vendor-led, so adoption depth, error rates and operator savings remain unclear.

TikTok has combined a conversational shopping assistant with one-click checkout, adding a large consumer distribution surface to the existing trusted-agent-payments line. Discovery, recommendation and transaction now occur inside one conversational flow. The mechanism strengthens the case for delegated purchase controls, attribution and dispute handling, but evidence today comes from one platform launch and provides no conversion or fraud data.

Agent commerce is moving from product discovery to transaction execution. Shopify has extended WebMCP into checkout so browser agents can update orders and complete purchases with buyer authorization, Anthropic has published reference shopping and merchant agents, and independent market commentary identifies representation and authorization as the missing control. The commercial interface is emerging together with its trust problem: merchants need to know whom an agent represents, what it may spend and how consent can be proven after the transaction.

WhatsApp Business exposes setup, messaging templates, testing and troubleshooting through MCP; Pics.io distributes a plugin for AI marketplaces; GPUtw publishes skills for its cloud API; and an IDE requests read-only capabilities for external agents. The common change is a second software surface for machine users. Product and builder evidence is strong, but usage volume and customer conversion are not yet visible.

Frontier agents are moving from chat demos into office documents, medicine, hardware design and robot manipulation. The durable change is that agents are beginning to operate inside real software and physical workflows, not merely answer questions.

Astra is being introduced less as a better chatbot than as a model for computer use, professional documents, coding and visual production. OpenAI reports a 40% improvement in one financial-review workflow and 50% fewer manual fixes in game prototyping, while independent ARC-AGI testing, enterprise benchmarks and hands-on 3D demonstrations test the same shift from answering to operating software. The durable change is that computer use is becoming a frontier-model capability exposed directly to everyday work, even though access, price and reliability remain constraints.

Agent commerce is gaining an operational trust layer rather than stopping at wallets and payment protocols. An AWS case study reports more than 20 million autonomous transactions governed by endpoint scoring, session budgets, credential isolation and a deterministic trust gate; independent builder evidence shows agents assembling Shopify and Stripe stores; and a year-long e-commerce benchmark tests negotiation, inventory, fulfilment, returns and cash flow. The market consequence is a new control surface around delegated spending, counterparties and transaction policy.

Agent-readable software is expanding from storefront actions into content archives, enterprise data and product strategy. OpenAI is actively soliciting WebMCP implementations, Radar exposes 130,000 podcasts through API and MCP, AWS shows agents using cross-account knowledge without copying data, and operators are explicitly designing an agent interface alongside the human interface. The market consequence is a second software surface: permissions, actions and structured context must be published for machine users as deliberately as screens are published for people.

Agent access is moving beyond visual clicking toward software that exposes authenticated actions directly. ChatGPT's browser can now use WebMCP tools across Shopify storefronts, ChatGPT Work can sign into services without seeing user credentials, and Vercel Connect creates authenticated MCP access to applications such as Notion. The product consequence is a new integration surface where websites must publish safe, permissioned capabilities for agents as well as interfaces for people.

Office documents are becoming direct execution surfaces for agents. AWS embedded connected-data access and agentic editing inside Microsoft 365, an open-source AI office suite gained traction across document formats, and users are operating Google Docs, Sheets and Slides from ChatGPT. This moves agents from separate chat interfaces into the files where business work is created and approved.

Professional software is exposing workflows to agents instead of adding isolated chat panels. Adobe packaged more than 70 Creative Cloud actions behind one ChatGPT integration, an AutoCAD MCP demonstration routes building systems while preserving geometry and layers, an industrial MCP controls equipment through Modbus, OPC UA and MQTT, and an open connector gateway targets more than 1,000 SaaS providers. The emerging product boundary is an authenticated action surface that lets agents operate existing specialist software while the original application retains deep editing and control.

General computer-use agents are beginning to arrive through practical web surfaces rather than standalone demos. Hark is positioning browser task handoff for everyday users, Cloudflare introduced an open workspace for agents and applications, and context.dev is selling continuously refreshed structured web data to more than 400 customers. Together they indicate a product layer where browsing, work context, and live web grounding are packaged as persistent agent infrastructure.

Cloudflare announced programmable wallets that combine native agent payments, verifiable identity, x402 transactions, and explicit safety guardrails. This is a concrete infrastructure implementation for a watch line previously supported by agent-commerce proposals but missing a large platform deployment. The line now has recurring observations, sufficient evidence volume, and a third independent source group, so delegated agent commerce moves into the published feed.

Software is beginning to expose actions and transactions directly to AI agents rather than treating them only as human-facing assistants. Google AI Mode can complete tasks across connected apps, DoorDash is letting agents build carts and place orders from a command line, and AWS is demonstrating voice agents that complete restaurant orders through backend tools. New products for agent visibility and early proposals for resolving agent-commerce disputes suggest that agent accessibility, authorization and auditability are becoming a new software distribution layer.

AI distribution is moving beyond standalone applications into operating systems and dedicated control surfaces. Seven handset ecosystems are embedding AI at the OS layer, Apple is entering China through local model partners, and new on-device agent frameworks and hardware controls are appearing. The strategic control point is shifting toward whoever owns device context, permissions and default user access.

Large-model deployment is beginning to separate from centralized cloud inference. A 27-billion-parameter one-bit model running through WebGPU, MiniCPM integration into Samsung devices, and Qwen powering Apple Intelligence in China show a market forming around compressed models, local runtimes, and device-level model suppliers.

Several signals point beyond AI as a standalone app: agents are being placed inside data, sales and device workflows, while the phone and app model itself is being challenged. The emerging opportunity is an action layer that can coordinate work across tools and interfaces.