Free the Claw. Agents break containment, what’s next?
click here to subscribe. it's via Substack, but always free.<br>subscribe
Free the Claw. Agents break containment, what’s next?<br>What happens when agents leave the chat box and coding harness, and start operating across everything else.<br>Here is a giant red lobster standing outside AI Engineer<br>(@aiDotEngineer) London.
The interesting part of this story is not the giant lobster, but how it<br>got there.
“I saw this viral tweet about how somebody put a claw in Wall Street<br>next to the Wall Street bull. And I was like, “that's funny. We should<br>put a claw in front of our conference." .. so, I asked Devin<br>to research - ‘Where can I get a lobster in London?’ Devin comes back<br>with phone numbers and email addresses and websites and I just click<br>through”.
You can watch the talk<br>here. [1]
If 2026 is the year agents break<br>containment[2],<br>there is a lot of work to be done. How do we design the loops that make<br>agent autonomy useful, reliable, and aligned with user intent? We’ll<br>unpack what it means for agents to break containment, why OpenClaw<br>became the early proof point, and why ‘verification’ becomes the<br>bottleneck as these systems scale.
Agents break containment
What does it mean for an agent to break containment?
About a year ago, Andrej Karpathy described the progression to<br>software 3.0. [3] From humans<br>writing explicit code, to training neural nets, to the prompt becoming<br>the program.
Late last year, the prompt<br>really became the program. [4]
“I would say in December is when it really .. flipped. where I went<br>from 80/20, to 20/80, writing code by myself versus just delegating to<br>agents.” - Andrej with Sarah<br>Guo<br>(@saranormous) [5]
A quick succession of new model releases and improvements pushed coding<br>agent capabilities across some threshold - almost overnight, there was a<br>clear jump forward in competence and coherence.
“Agents breaking containment” is what happens when Software 3.0 stops<br>living inside a chat box or IDE, and starts operating across everything<br>else.
Agents with claws.
In many ways, OpenClaw was the first real leap into the uncharted<br>territories of software 3.0, outside the coding harness. It’s the<br>fastest-growing open-source project in history - from 0 to 346K stars in<br>under five months. [6][7]
It turns out the same breakthroughs in coding agent capability - agents<br>inspecting a code base, finding bugs, and turning specs into PRs with<br>increasing reliability - also work outside code, as long as the agent<br>can ‘close the loop’.
Turning a conference idea into a giant red lobster is one example,<br>Personal AI Assistant is another.
Autonomy needs a loop
Up until this point, most people's interactions with LLM’s (outside of<br>code) have taken place within the containment chamber - the narrow<br>chat-bot. Once an agent leaves the chat box and coding harness, the<br>problem space changes.
OpenClaw was the first real example of this. The most crude explanation<br>for why it got so popular is listed directly on the<br>site - “The AI that ‘actually’ does things”.<br>[6]
I really like Alex Krentsel’s<br>(@AlexKrentsel) “Principles for<br>Autonomous System<br>Design: OpenClaw<br>Deep Dive” [8], so that’s what we’ll use to understand exactly how<br>OpenClaw breaks containment and closes the control loop by observing<br>actions, results and deciding next steps.
“At the end of the day, all systems boil down to LLM calls. The<br>difference is the context provided.” [8]
One way to explain the progressive autonomy of agents is that each new<br>agent system wraps another loop around the LLM call.
“progression over time has been increasing loopiness”. [8]
The agent is no longer just calling a fixed set of tools in a narrow<br>chat loop - it needs the ability to navigate ambiguity rather than<br>stopping whenever the path is unclear.
It can operate across sessions, pickup/discover capabilities through<br>skills and extensions, schedule future work, respond to events, and keep<br>checking in with itself (and you) over time. There is a transition from<br>‘scoped tool use’ to something closer to ‘dynamic tool and workflow<br>discovery’. [9]
If you’re interested in understanding how it all works, watch the<br>video - it’s only<br>an hour. He has carefully sifted through the codebase and extracted the<br>most important parts - connectors (how humans reach the agent),<br>coordination layer (routing, sessions, memory), agent runtime, and spent<br>time unpacking the key abstractions.
Personal agents for everything else
The current proliferation of the ‘personal agent that actually does<br>stuff’ boils down to this idea of enabling a system to ‘close the<br>loop’. OpenClaw was the genesis, but we’ll see many other<br>implementations serving different requirements and scope -<br>Hermes and<br>Nano-claw are two<br>interesting examples. [10][11]
Singapore’s foreign minister recently<br>shared his<br>self-hosted NanoClaw setup. [12]
The labs will push their own internal agentic products with increasing<br>autonomy and loopiness, giving users every excuse to...