In early 2025, we were seeing an explosion of AI agents:
- in the product announcements from major platforms (OpenAI, Anthropic, Langchain, etc.)
- in no-code ecosystems like Make, n8n or Zapier
It was the big trend of 2025. And it also means adding a layer of intelligence to automations.
That is the topic of this article: behind all the buzz, which agents are actually in production? Where do we really stand? What are the impacts for us?
From deterministic to probabilistic
Where automation used to be limited to rigid, linear scenarios, we are entering a more flexible, more adaptive era. Tools are becoming able to make decisions in context, handle exceptions, and even run entire processes autonomously.
In April 2025, for instance, Make added its AI agents feature to scenarios, in beta. At that stage it was a bit early (v1), but it signals a real paradigm shift: no more fully deterministic logic, making way for smarter, contextual automation that can adapt to the ups and downs of the real world.
Here is a preview in this video:
Concretely, it means less maintenance thanks to adaptive agents, better coordination between modules and scenarios, and optimized contextual decisions.
The reality of the market today
Despite all the excitement around AI agents, the reality on the ground is very different: very few agents are actually in production inside companies.
Because an AI agent is not just a module you plug in (contrary to what you might see).
You need a structured process, well-defined flows, a coherent data architecture, and in-house AI skills. Without that, it is hard to move from the idea to concrete implementation. And most companies are not at that stage.
Today, I mainly see 2 broad options:
- The turnkey agent: deployed on classic, recurring business use cases: customer support with Decagon, Sierra, Maven AGI. Sales prospecting with 11x, Artisan, … Knowledge management: Glean, Dust, Unleash, … Legal with Harvey, etc.
→ These are agents where the last mile remains to be done: the implementation part and the business logic.
On that note, Decagon (support) is really interesting with its concept of turning SOPs (Standard Operating Procedures) into AOPs (Agent Operating Procedures).
- The custom-built AI agent: much more flexible, but also more complex. It requires involving the technical team together with the business team (or going through an external partner), with an iterative approach and, above all, a maintenance plan. It is rarely feasible for small organizations… unless they have a strong tech culture.
For now, moving to production remains limited, but the situation will open up as the ecosystem takes shape.
For example, the accessibility of no-code tools and the arrival of simplified orchestration solutions will let many small businesses build their own AI agent easily.
Personally, I expect to see many micro AI agents emerge in small and mid-sized businesses.
For example, n8n already lets you get an AI agent skeleton in 1 click today
Then let the agent fill in the variables when using a third-party tool:
Overall, we delegate to the agent the intelligence needed to run the automation → it figures out how to reach the goal we give it in plain text (in the prompt).
But be careful not to use AI where a classic automation would be enough. We still reach for AI agents too often just because they are trendy. In most cases, we lose efficiency, complicate maintenance, and dilute the real value.
October 2026 update: since I wrote this, agents in production have multiplied without becoming the norm. In November 2025, McKinsey found that 23% of the companies it surveyed were scaling an agentic AI system in at least one function, and 62% were at least experimenting with agents. In June 2025, Gartner predicted that over 40% of agentic AI projects would be cancelled by the end of 2027, due to escalating costs, unclear business value or inadequate risk controls.
The ecosystem is taking shape
The AI agents ecosystem is taking shape… and it is moving very fast.
First, MCP (Model Context Protocol) lets an agent connect to third-party tools:
→ Claude can query Perplexity, update Notion, or trigger an action in your CRM.
These protocols industrialize the range of actions AI agents can take, as I explained in my March 2025 newsletter issue on MCP.
On the other side, it continues with the launch of A2A (Agent-to-Agent), a protocol proposed by Google to let agents collaborate with each other, even if they were never designed to. It is a small revolution.
Before, each agent worked in a silo, within its own scope. Today, they can discover each other, split up tasks, cooperate in real time, without sharing resources or memory.
In short, A2A brings a collective brain to AI agents.
It opens the way to multi-agent workflows, specialized by function or by tool, systems that are more robust, scalable, and adaptive.
5 best practices for AI agents
Before moving on to the grey areas of the AI agents ecosystem, I wanted to share 5 best practices to set up an effective AI agent.
Before building, take the time to map out the framework, the key steps, the critical decisions, and the points AI could optimize.
You can use FigJam to do this step.
Avoid catch-all agents. Start simple with one goal and 1 to 3 tools maximum. Then split it up if needed into specialized sub-agents.
A good system prompt makes all the difference: explicitly stating the available tools, how to use them, the parameters to respect, and the limits and rules to follow…
RAG (retrieval-augmented generation) is faster and clearer than an AI agent in many cases.
The more freedom the agent has, the fuzzier debugging becomes. In a classic workflow, every step is traceable. With an agent? You have to guess what hallucinated. On some scenarios that can be problematic, otherwise you have to accept this extra layer of complexity on the debugging side.
The grey areas: toward an agentic economy?
What if, in the coming years, a company's workforce were no longer human?
That is roughly where AI agents could take us… And 2025 looked set to be that pivotal year.
This shift does not only affect operations: it also upends business models.
No more pricing per user or per SaaS seat. We now talk about value created:
→ cost per task solved, per customer reply, per piece of content generated, per action executed.
A glimpse of the new pricing models arriving with AI agents:
It is the emergence of a new workforce:
- The traditional era (employees, offices, payslips)
- The digital era (freelancers, remote, SaaS)
- The agentic era (automated, non-human)
But this shift raises as many promises as grey areas
- What becomes of a SaaS designed for human users… when those users become AI agents? → I covered this in a section of my newsletter on the end of SaaS.
- How do you invoice or tax a sales AI that generates 2M€ in revenue with no company registration, bank account, or payslip?
- Who is responsible if an AI agent crashes a system or makes a bad decision?
- Can we imagine A2A (agent-to-agent) negotiations between a sales AI and a buyer AI?
- Does an AI agent have “labor rights” if we start treating it like an ultra-efficient autonomous worker?
Today, the AI agent is a tool.
Tomorrow, it is a full-fledged economic player, and nobody really knows how to fit it into the current system…
An infrastructure still to be invented
For this agentic workforce to become viable at scale, all the surrounding infrastructure is still missing:
An equivalent of payroll (task tracking, pay based on output)
An agent identity (profile, history, traceability)
A system for compliance (regulatory, tax, legal)
The ability to define the intellectual property produced by agents
The change is already here. But the framework itself is still to be written…
As with every industrial revolution, it is not the technology that is missing, but rather the rules to govern it.
For now, we can already prepare for this future by laying the groundwork: measuring the value an agent creates, tracking its real efficiency, thinking in “mixed teams” (humans + AI)… and staying curious in the face of this transformation that is only just beginning. To move from curiosity to a first agent in production, we start with an AI audit.
