By Pascale Caron — based on the “AI Agent Bible” report, CB Insights, 2025
At the dawn of 2026, artificial intelligence is entering a new era. After the age of assisted chatbots, “AI agents” are now emerging as autonomous entities, capable of acting, reasoning, planning, and interacting with complex environments. According to the “AI Agent Bible” report published by CB Insights in 2025, over 500 startups have been founded since 2023, structuring a booming market.
This shift marks a paradigm change: AI no longer merely assists—it executes. In many cases, it anticipates. Cloud giants, software vendors, and young startups are converging toward the same goal. They aim to create intelligent agents capable of transforming workflows, reducing interaction costs, automating low-value tasks, and strengthening operational resilience.
From Guarded Agents to Full Autonomy
Today, the majority of agents operate in structured environments. Guardrails frame their actions. They execute tasks within structured workflows, with limited decision-making capacity. But CB Insights’ projections anticipate rapid evolution: by 2026, truly autonomous agents—acting without human intervention—are expected to emerge in critical environments.
This prospect is based on the rise of more robust foundation language models (LLMs), but also on the development of new interface forms. Voice applications, AI-driven browsers, or even “AI-native workspaces,” built from the ground up around the agent.
A Strategic Challenge for Businesses
Pressure is mounting on executive leadership. Integrating AI agents is no longer an experimental luxury: it’s a strategic necessity. CB Insights notes that in financial services, healthcare, and legal sectors, agents already assist clinical decisions, assess risks, and draft legal briefs.
25% of startups founded since 2023 have already reached commercial distribution levels previously unattainable in under five years. An unprecedented phenomenon according to the “commercial maturity” indicators developed by CB Insights. The leverage effect is clear: a well-resourced startup with data can disrupt a market in months.
The Battle of Infrastructures and Standards
Behind the excitement, a silent war is unfolding around infrastructure. Three cloud giants—Amazon, Microsoft, and Google—are competing for dominance over interoperability standards. Anthropic’s Model Context Protocol (MCP), Google’s A2A protocol, and IBM’s inter-agent communication standard aim to standardize exchanges between agents and third-party tools.
This battle of standards is significant: it prefigures the future Internet of agents. One where intelligent entities will need to collaborate, authenticate, exchange sensitive data, and negotiate complex actions among themselves. Victory will go to those who master not the models, but the orchestration, coordination, and governance layers.
Three Innovation Fronts: Voice, Payments, Security
The report identifies three segments to watch. First battleground: voice. In 2025, startups specializing in voice interface development raised $400 million. Meta, for example, acquired PlayAI and WaveForms AI, signaling a shift toward voice interface as the primary vector for human-machine interaction.
Second front: payments. The challenge here is to enable agents to conduct secure transactions in real time. Stripe, in partnership with OpenAI, launched the “Agentic Commerce Protocol,” which defines a standard for automated payments. A decisive step toward transactional autonomy for agents.
Third axis: cybersecurity. The proliferation of agents opens new attack surfaces. Players like Zenity, WitnessAI, or TrojAI are developing tools to monitor behaviors, assess reliability, generate synthetic users for testing purposes, and produce auditable reports.
Verticalization: Toward Industry-Specific Agents
Beyond horizontal applications (customer support, software development, HR), agents are now tackling vertical industries. In financial services, companies like Boosted.ai or Wokelo assist financial research. In healthcare, Hippocratic AI structures end-to-end workflows, from virtual triage to billing cycle management.
In manufacturing, the agent becomes an operator. The company Composabl offers agents capable of directly controlling industrial equipment. Palantir is also very active in this domain. The objective: reduce dependence on human intervention in complex environments.
The Agent Economy: Emerging Business Models
Monetizing agents is becoming a strategic issue. According to CB Insights, agents specialized in software development—like Cursor ($500M ARR) or Replit ($150M)—dominate the revenue race. They display impressive returns, averaging $1.4 million in revenue per employee.
Conversely, customer support agents, while less profitable in the short term, benefit from higher valuation multiples: 219x on average. This reflects investor expectations regarding their potential for mass replacement of human teams.
The question of profitability remains sensitive, however. CB Insights anticipates margin compression, linked to exploding computational costs of models with high reasoning capacity. Salesforce, for example, revised the pricing of its “Agentforce” agent in May 2025, shifting from a per-conversation model to a usage credit system.
Startup Ecosystem: The Structuring Role of Accelerators
Y Combinator plays a pioneering role in this ecosystem. The Spring 2025 cohort shows that over half of selected startups are developing AI agents. Among them are players specializing in code testing (Docket, Cubic), legacy system integration via browser agents (Kaizen), or back-office agents targeting accounting, HR, or reporting functions.
Some startups go further and are beginning to tackle autonomous research. Bramante Biologics or Scalar Field, for example, aim to automate documentary exploration in healthcare or finance. Their models could eventually replace certain documentary research or primary analysis positions.
New White Spaces: Governance, Audit, Marketplaces
CB Insights identifies three high-potential “white spaces.” First, marketplaces. AWS launched its own in July 2025. Players like Agent.ai seek to differentiate through personalization and discovery of specialized agents.
Next, cost management. Larridin, backed by Andreessen Horowitz, enables companies to track their AI spending and return on investment from deployed agents.
Finally, governance. The question of traceability, explainability, and regulatory compliance is becoming central. Players like LangWatch, Coval, or Traceloop are developing solutions for evaluating agent behaviors.
The Agent as Supertool: A New Grammar of Work
As Michael Mignano (Lightspeed) emphasizes, we are moving beyond the era of the passive copilot. The agent becomes a “supertool,” no longer simply responding but shaping the organization, anticipating needs, interacting with systems, and structuring how a company learns and adapts.
In this new context, successful companies will be those that can connect specialized agents to specific data, within a coherent, manageable, auditable ecosystem.
Major Challenge: Contextual Memory
Agent memory remains fragmented today. Each interaction is often treated in isolation. CB Insights emphasizes the importance of “context engineering”: unifying signals from email, CRM, support systems, and SaaS usage, so the agent has the right context at the right time.
Startups like Letta, LlamaIndex, or Zep AI are seeking to build these contextual layers. The challenge: creating interoperable, continuous, transparent, and governable memory. Memory that is not merely technical, but organizational.
Systemic Risks and Regulatory Challenges
Agent autonomization raises ethical and systemic questions. How can we ensure that an agent left alone in a critical system won’t induce erratic behavior? How do we audit a decision made by a cascade of agents?
CB Insights does not minimize these risks. The report suggests that developing surveillance layers—observability, auditability, explainability—must precede any widespread deployment of highly autonomous agents. Regulation will need to adapt to this new algorithmic layer of decision-making.
Toward Augmented Organizational Intelligence
The AI agent is not just a tool. It redefines the organization. In pioneering companies, it evolves into a cognitive partner. It learns, remembers, analyzes, acts, cooperates. It becomes a catalyst for adaptation, a vector of innovation, and a factor in cultural transformation.
The successful company will not be the one that adopts generic agents, but the one that can orchestrate a network of custom agents, built around its industries, flows, and data.
An Economy Shaped by Agents
As Manlio Carrelli (CEO of CB Insights) concludes: “The competitive advantage does not lie in AI, but in its capacity to be operationalized.” The future will not be determined by models, but by their insertion into business architectures.
AI agents are already at work. They are transforming how we code, recruit, respond, and decide. Tomorrow, they will reshape value chains. Those who understand the rules of this new agentic economy will write its future.
Reference
CB Insights (2025). “AI Agent Bible: The ultimate guide to agent disruption.”
Available via CB Insights Business Graph. Published in March 2025.
