Weekly Roundup: Tencent's $2bn Agent Bet, Identity Security for AI Workforces, and Agent Funding Hits $1.8bn

This week the agentic AI market crossed another threshold. July funding reached $1.8 billion across 12-plus deals, Tencent signalled a strategic move worth over $2 billion into agent startup Manus, and identity security platforms began treating AI agents as first-class users. The shift from "can we build agents?" to "how do we govern, secure, and fund them at scale?" is now visible across every layer of the stack. Here is what happened and why it matters for enterprise buyers and suppliers.

Capital flows: Tencent's $2bn signal and the maturing agent funding market

Tencent is reportedly eyeing an investment of over $2 billion in AI startup Manus, in what would be one of the largest single bets on an agent-native company to date. The move signals that major technology platforms see agent infrastructure as a strategic asset rather than a venture experiment.

At the same time, Nous Research closed a $75 million Series B at a $1.5 billion valuation, led by Robot Ventures, validating the market for open-source agent development. And Bespoke Labs raised $31.75 million in Series A funding for AI agent training, underscoring that the tools for building agents are becoming a distinct and well-capitalised subcategory.

According to AI Funding data, July's 12-plus deals totalled $1.8 billion across the agent category, a 35 per cent increase over June. Significantly, 62 per cent of those deals were Series B or later, averaging $150 million each with median ARR of $25 million-plus. The speculative seed phase is giving way to revenue-backed scale-ups. Sequoia Capital led four deals, including two above $100 million, and 42 per cent of deals closed outside Silicon Valley, with London, Tel Aviv, and Paris emerging as secondary hubs.

Why it matters. The funding market is separating into two tracks: early-stage exploration, and late-stage conviction backed by hard revenue metrics. For enterprise buyers, this means the agent suppliers likely to survive and scale are those with $25 million-plus in ARR and demonstrated enterprise contract performance, not demo-stage startups. For suppliers, the message is clear: series B capital is available, but only with evidence of production revenue and measurable ROI in vertical-specific use cases.

Identity and security: AI agents become first-class users

Okta and Google Cloud announced a partnership to extend identity security to AI-powered workforces, treating AI agents, users, and devices within a unified security framework. Okta separately held a briefing outlining its platform expansion to treat AI agents as first-class identities, adding centralised discovery, registration, entitlements, access reviews, and enforcement controls. The company frames this as a "non-human identity" category that includes machine, bot, and agentic AI accounts.

Entrust launched its Agentic AI Trust Accelerator, a tool designed specifically to help enterprises move AI agents from pilot phases into full production by streamlining identity, access, and compliance workflows. BMC Software brought governed AI agents to enterprise workflows and mainframe operations, adding secure access controls and actionable intelligence retrieval for agentic assistants working across core business systems.

On the threat side, Straiker Research reported that 36 per cent of successful AI coding agent attacks result in remote code execution, while Zenity highlighted the unique security challenges posed by coding agents, arguing that traditional application security approaches do not adequately cover agent behaviour. Cloudflare introduced an AI-powered bot defence engine, describing the shift as preparation for an internet increasingly dominated by AI agents.

Why it matters. Agent identity is becoming a distinct security discipline. When an AI agent accesses your CRM, sends emails on behalf of an employee, or invokes financial APIs, it needs credentials, entitlements, and audit trails, just like a human user. Buyers should be asking every agent supplier how their product handles identity provisioning, access scoping, and session logging. Suppliers that integrate natively with identity platforms like Okta will have a measurable advantage over those treating agent access as an afterthought.

Enterprise agents go operational: from pilot to governed production

Fujitsu announced a field trial deploying an AI agent alongside store managers at AEON Food Style in Japan, moving agentic AI into physical retail operations rather than digital-only workflows. The agent works with human managers to optimise inventory, staffing, and operational decisions in real time.

Reliance Global Group launched a proprietary AI agent for secure browser automation in regulated insurance back offices, demonstrating that the most challenging deployment environments (highly regulated, legacy systems, sensitive data) are now accessible to agentic automation with the right governance controls.

ActiveCampaign released Active Intelligence 2.8, an update that learns brand identities to draft marketing assets and automated sequences, bringing agentic content creation into the marketing operations mainstream.

Why it matters. Enterprise agent deployment is diversifying across sectors. Retail, insurance, and marketing operations are all now running live agent deployments, not pilot programmes. The common thread across Fujitsu, Reliance, and ActiveCampaign is governed operation: each deployment includes explicit controls for data access, human oversight, and auditability. For buyers, this means governed production is now the baseline expectation, not a premium feature. For suppliers, the differentiation is shifting from "our agent can do X" to "our agent can do X securely, auditably, and within your existing compliance framework."

Infrastructure and developer tools: the agent stack deepens

Microsoft released a Go implementation of its Agent Framework, providing a foundation for building and orchestrating production-grade multi-agent workflows. Google integrated agent-building capabilities into its open-source Genkit framework, enabling developers to create conversational and multi-agent AI applications without switching toolchains. Meta Muse Spark 1.1 brought agentic AI, coding, and multimodal reasoning to developers via the Meta Model API.

Separately, Codex announced encryption for multi-agent prompts, a capability that addresses data confidentiality during agent-to-agent communication but raises questions about local auditability for enterprise compliance teams.

Why it matters. The infrastructure conversation is moving from "which model?" to "which stack?" Go, Genkit, and Meta's API represent three different approaches to the same problem: how to build reliable, observable agent systems that integrate with existing enterprise toolchains. For buyers, this proliferation means procurement teams need to evaluate agent frameworks on the same criteria they apply to application platforms: ecosystem support, integration depth, security posture, and long-term maintainability. For suppliers, framework choice is becoming a strategic decision that shapes addressable market, hiring, and partnership strategy.

Market intelligence: enterprise adoption at 25 per cent, revenue proof arrives

A new report by FirstPageSage combining data from 16,000 companies found that enterprise organisations now lead agentic AI adoption at 25 per cent, with mid-market and SMB segments catching up quickly. The report describes the agentic AI market as "AI's next $1 trillion opportunity."

Analysis from MarketingProfs reinforced that enterprises are becoming less willing to build everything around a single model provider. They are instead separating models, agent frameworks, gateways, data systems, and sandboxes so components can be changed independently. Price and production performance are driving increased adoption of Gemini and open models alongside OpenAI and Anthropic systems.

The AI Funding analysis identified three structural drivers behind the funding acceleration: enterprise revenue validation (coding agents at $50 million-plus ARR with 150 per cent-plus net revenue retention), model commoditisation shifting defensibility to agent orchestration layers, and agentic software replacing point solutions. A single coding agent replaces linters, code review tools, documentation generators, and junior developer hiring. An enterprise research agent replaces subscriptions to multiple data providers and analyst headcount.

Why it matters. The economic logic for agentic AI is crystallising. When one agent platform replaces five to ten point solutions, the business case writes itself. For buyers, the practical question is now about vendor consolidation strategy: which agent platforms are broad enough to replace the tools you currently license separately? For suppliers, the consolidation dynamic rewards platforms with broad integration ecosystems and punishes narrow point solutions that cannot demonstrate a measurable return on the total cost of ownership.

This was the week that agentic AI funding, identity security, and enterprise governance all converged on the same signal: the market is no longer asking whether agents work. It is asking how to deploy them safely, fund them at scale, and integrate them into the same governance frameworks that protect every other part of the enterprise.