root@fr:~$ man agentic
The agentic era has begun.
In two years, AI moved from conversation to action. Here is what changed, what an agent really is, and why an agent without governance never leaves the prototype.
Put simply
What is an agent?
An agent is software you hand a task, not a question. You give it a goal: it reads the relevant documents, uses your tools, works step by step and presents the result for your approval. A conversational assistant answers you; an agent hands you the work, done.
Example: "prepare the summary of this forty-exhibit file". The agent reads the exhibits, writes the summary, cites the page behind every statement, and waits for your review.
The shift
From answering to acting.
In 2023 and 2024, generative AI answered: you asked, it generated text or content, under heavy human oversight. Since 2025, agentic AI observes, plans and acts: it chains multiple steps, calls tools, draws on memory, self-corrects and pursues a goal with less supervision. Forbes put it plainly in late 2025: step aside chatbots, agents are the next stage.
2023 - 2024
Chatbots and generative AI
Reactive: chat and generate
- You ask, the model answers
- Conversation and content production
- One exchange at a time, no lasting memory
- Heavy human oversight at each step
2025 - 2026
Agentic AI
Autonomous: observe, plan, act
- The agent observes, plans, then executes
- Multi-step workflows
- Tool use, memory and self-correction
- Goal pursuit with lighter supervision
The same question, two postures: one waits for an instruction, the other pursues a goal.
Three words, no more
Agent, agentic, harness.
The vocabulary of the agentic era fits in three terms. Confusing them is confusing an engine, a behaviour and the car around them.
An agent
Software, often LLM-based, that pursues a goal: it perceives, plans, calls tools and APIs, acts, then iterates while drawing on memory.
Agentic
The autonomous behaviour itself: deciding, planning multiple steps, executing and adapting with limited human input, as opposed to a purely reactive chatbot.
A harness
The infrastructure wrapping the model: orchestration, memory, tool interfaces, guardrails and runtime. It is what turns a raw model into a reliable, auditable agent.
Agent = Model + Harness
The formula is used by LangChain, by Martin Fowler under the name harness engineering, and by Salesforce. The model brings raw power; the harness brings reliability, memory and control. Without a harness, a model stays a demo.
The maturity gap
Most pilots never reach production.
Adoption is moving forward, but very unevenly. The jump from pilot to production is the wall where most projects stop.
Explore
Roughly 30% of organizations are exploring agentic AI, per Deloitte. They test, scope and learn.
Pilot
Roughly 38% are piloting, per Deloitte. A proof of concept runs, but outside real production constraints.
Production
Only 11 to 14% reach production, per Deloitte. This is where projects fall: reliability, oversight, governance.
Gartner expects about 40% of enterprise applications to embed task-specific AI agents by the end of 2026, up from under 5% in 2025. The same analysis places agentic AI at the peak of inflated expectations. A forecast is not a fact: it is a reason to aim straight for production.
It has been our requirement from day one: production-ready, not prototypes. Reliability, oversight and graceful degradation, from the start.
The open harness wave
Agentic AI already runs on open, self-hostable models.
Two open-source projects show how mature self-hostable harnesses have become. We describe them factually, with their strengths and their weaknesses.
OpenClaw
Open-source, self-hostableOpen-source autonomous agent, originally released in November 2025 as Warelay. It runs real tasks via language models accessible through messaging apps. Created by developer Peter Steinberger, who joined OpenAI in early 2026, it is now stewarded by the non-profit OpenClaw Foundation. It went viral, passing 100,000 GitHub stars. It has documented security weaknesses.
Hermes
Open-source (MIT), self-improvingSelf-improving agent harness released under the MIT licence by Nous Research, around February 2026. It keeps persistent memory across sessions, reuses Skill Documents, manages subagents and runs local or in the cloud.
Both projects prove one thing: agentic AI can run on your own infrastructure, on open models. Sovereignty is therefore possible. But their security gaps, which confine them to experimentation, prove the other half of the same truth: you need a governed, sovereign harness. The root Workspace is exactly that, a sovereign, governed agent harness. An agent can work there overnight with no one at the keyboard; even then, nothing is written into your tools without human sign-off. Mycroft is one of the agents that run on it.
An agent is worth its harness.
The agentic era rewards those who master the infrastructure around the model, not just the model. Build on a sovereign, governed, auditable harness.