For the past couple months, my partner Hamid and I have been running an experiment. We installed NanoClaw, a variant of the hugely popular agent system OpenClaw. Despite the known risks of giving such software access to our critical business information — along with an ability to act on our behalf — why did we do it? For sure, there was some FOMO, but we also genuinely wanted to find out how much an innovation consultancy could be transformed, or what might be possible, if we gave agentic software the ability to work with us. And for organizations and enterprises we supported, we wanted to be able to analyze the pros and cons, and provide insight.
So far, a few months in, the impact for us is measurable and positive. And not in the ways we necessarily expected. We’ve optimized some existing workflows, and new ones were automated. But what my partner and I have stumbled upon feels like a new way of working together. It has turned the previously two-way dialogue between us two co-founders into a three-way one, between us and an “agent” (actually several of them), which remembers everything we talk about and proactively suggests new things we can do to move our business forward. There are devils in the details, of course. As the kids say, “It’s complicated.”
A Quick Primer on ‘Claws’
The breakthrough insight that OpenClaw creator Peter Steinberger stumbled upon was that the limited success of agents to date was not the fault of the models (Claude, ChatGPT, Gemini, etc.), but rather a missing bit of software, called an “agentic harness”; i.e., a user-controlled system that would run in a user’s own local environment (like, on a spare Mac Mini computer). This “harness” would act as a wrapper around the big frontier models, holding memory, and executing programs (called “skills”) on schedules, and even creating new skills on the fly. Steinberger also standardized the system in such a way that these skills could be easily shared, enabling a global skills marketplace and sparking network effects. And as we’ll see below, all of this enabled a target-rich environment for hackers and malicious actors.
OpenClaw allowed teams to automate ongoing delegated work, and it could improve its own workflows.
OpenClaw allowed teams to automate ongoing delegated work, and it could improve its own workflows. The processes operated in the user’s own controlled environment (not in someone else’s cloud), and on a cadence (called a “heartbeat”) that allows agents to run extended flows, surfacing occasionally when they have questions, as they make dozens if not hundreds of decisions on a team’s behalf.
Two other contributing factors to OpenClaw’s success were user interface breakthroughs. First was mobility. Gone were the days of being tied to another vendor’s chat screen. Instead, you can engage OpenClaw through your own mobile tool such as Slack, WhatsApp, or Discord. It’s sort of like leaving a voicemail message. You can speak to WhatsApp on the way to work, saying, “Take last week’s customer interviews from project X and build three new pilot features in our coming product. Launch them on our test site. Push the code to our GitHub.” By the time you arrive at the office, you pick up the result and your team is already notified of the progress.
The second UI breakthrough was the installation process. Agent software, as you can imagine, is geeky stuff, involving command lines, path names, obscure libraries, and packages. Installing such tools in your own computer environment is not for the faint of heart. But anyone who has even started to tinker with Claude Code can say, “Hey Claude Code, I have downloaded OpenClaw, figure out how to install it and get it working.” And it just does.
The reason that magic works so well is that Steinberger designed the installation process for agents, not people. In other words, he put the installation details in files called Markdown, which Claude Code (or Codex, or Gemini) can read and then infer just what to do, no matter which environment the user is installing into. None of the traditional “download this binary for Mac OS,” “download this one for PC,” etc. The agent figures it all out, taking the necessary steps to make the system just run.
Choosing Your Claw
Our first job was deciding which claw to choose and deploy. There are, at the time of this publication, a dizzying number of so-called OpenClaw “variants,” easily discovered via GitHub searches. To whittle the list down, you might try a tool like ClawCharts, a public leaderboard that ranks the top options by star ratings, and shows their relative adoption momentum based on a composite metric of downloads, number of active contributors, etc. No surprise, OpenClaw is still far and away the leader. If you scan that list, they fall into two main groups: those optimized for sheer strength of agent orchestration (OpenClaw, Hermes Agent, Paperclip), and those designed for security with very minimalist runtimes and footprints (ZeroClaw, PicoClaw, and NanoClaw).
After doing our own research, we leaned toward the minimalist camp. My partner recommended NanoClaw (currently #7 on the charts). NanoClaw, as the name implies, has a smaller footprint (5,000 lines of code vs 250,000 for OpenClaw), so in theory a smaller surface area for security exploits. It was created by Gavriel Cohen and announced on January 31, 2026. His philosophy is explicitly security-first. Each agent is assumed to misbehave, so NanoClaw only allows each messaging group (WhatsApp, Telegram, Discord, etc.) to run in an isolated “container,” a lightweight virtual machine called Docker, completely separate from your OS. Only explicitly-mounted folders are visible to the agents (not your whole disk by default.)
Note for Enterprise Installations
Security concerns are the number one worry with this agentic harness technology. An open marketplace of claw skills is great — until a user naively installs one and injects a Trojan horse into the system that acts as spyware. Yes, it has happened. Claws are not standardized for the enterprise yet, but that is coming. The most notable move towards “enterprise-ing” such agents is Nvidia’s NemoClaw. Others (OpenAI included) are not far behind.
NVIDIA CEO Jensen Huang clearly sees the momentum behind [agents,] as well as the adoption barriers for enterprises around security.
NVIDIA CEO Jensen Huang clearly sees the momentum behind agentic harnesses, as well as the adoption barriers for enterprises around security. NemoClaw promises “Safer AI Agents and Assistants.” It’s also designed to sell NVIDIA hardware, since it runs on their chips. OpenAI hired OpenClaw creator Peter Steinberger, and has clearly prioritized security and enterprise hardening. IBM Research is tracking the spate of security issues. Their recent coverage noted that “approximately 15,000 vulnerabilities have been disclosed so far in 2026.”
What Our Team Has Learned
In just under two months, we’ve gone from the starting gate to a nine-channel (nine agents) operational architecture, which enables autonomous publishing, meeting intelligence, multi-project tracking, shared institutional memory, and a meta-skill that writes its own extensions. Our primary agent, La Force, is the one we spend the most time interacting with. In terms of sheer volume of activity, between installation on March 21st and the time of writing this article, we have had 664 documented exchanges with La Force, and hundreds more with its sibling agents.
All of this is happening for us in Slack. Each Slack channel is another agent and another persona. For example, one of them does meeting transcriptions and analysis. Its follow-on meeting analysis is phenomenal. “Curtis, you spoke way too often and didn’t let the interviewee go on long enough,” it has told me. Yes, it can be blunt if you ask it to.
Another dedicated agent for us involves building reports. We have very specific guardrails on look and feel, and what we want readers to experience. This is encoded in the skills, partly deterministic (code) partly probabilistic (agent/LLM). When you see reports like this or this, those came out of a specific agent channel called Collision Brief Reporter. It takes in a set of incoming signals, finds connections and themes, and puts the story together. This is automated publishing, where the human input is our curation process (which sources we pull from), the guardrails we give the agents, the limits we set, and so forth.
Our marketing channel agent, called Jedi Mind Tricks, has access to our Google Analytics data. It runs a weekly report, comparing numbers from prior week, and then makes suggestions on new messages to try. General Sun uses military strategy to pressure test our ideas.
Words to the Wise
Here are a few “first principles” for if/when you stand up an agent claw system for your own workgroup.
- Mindset Shift: From Chat to Flow. When we first started, it was unclear the value of the system until we moved from thinking of question -> answer to need -> workflow. For example, my partner said on a lark, “Create a skill that can detect a web tech stack just by reading the HTML and focusing on the header component.” It did it and then we made it a core skill. You’re not adopting a product when you use these tools. You are holding a product creation tool in your hands, one that can “productize” just about anything it can get its digital “claws” around.
- Prepare for Smart and Dumb: At their best, agents do just what you hoped they would each time. But inevitably, you will give the claw an edge case, and it will drop the ball, just like a college intern might. Treat it as such, and plan to keep adding specificity to agent skill instructions as you go. You don’t have to hand edit anything. When it makes a mistake, just ask it to update its skill to handle this edge case the next time.
- Single Purpose, Single Channel: Claw agents operate in communication channels (think Slack, WhatsApp, Discord, etc.) Create as many channels as you have specific innovation functions. For example, we segmented transcriptions into its own channel that we called MeetOS. Another channel is for marketing functions, another for tech research, etc. Along with each singular purpose, give each agent what Peter Steinberger calls its “soul,” essentially a persona. Our agent La Force is Enneagram Type 5, objective critic, sycophancy removed.
- Infer Then Encode: Demonstrate or prove out a workflow with an agent, then make it stronger. Consider an innovation workflow like “customer interviews → insights” with all the steps on that journey. Give the entire task to the agent. Prove it out. Then, try to migrate as many aspects of that agent workflow into “deterministic” steps as possible; as in, convert them to Python scripts that run exactly the same way each time. This can greatly reduce agent “improvisation.” Reserve your LLM token budget for only those steps in the workflow where fuzzy judgment is actually needed.
Conclusions: It’s Not Just About Productivity
Our NanoClaw system, nicknamed La Force, is extremely capable, but is not perfect by any means. According to some new research from a group of Oxford University researchers at Kings College, no such agentic system will ever be perfect. The gap between ideal predictable rule-like behavior and the messy reality of execution will persist; it’s Kurt Gödel’s Incompleteness Theorem applied to AI.
For sure, our claws occasionally space out, repeating a mistake we thought we had corrected earlier. It’s the sort of experience we’ve all had with generative AI, Ethan Mollick’s so-called “jagged edge.” At its best, our system runs production tasks at 100x speed, building complex reports and client deliverables such as our bi-weekly InnoLead enterprise AI report, which scans thousands of discrete signals to find items of relevance to innovation groups.
In its lesser moments, we babysit tasks from time to time, as noted in the “Prepare for Smart and Dumb” section above. The good news for us thus far, however, is that the former (clean execution) outweighs the latter (missteps) by about an 80-20 ratio. As my Dad the electrician used to say, “That’s good enough for government work.”
These agents are becoming our new colleagues, warts and all. Like our real-life flawed personal colleagues, they can make us laugh. When each agent has its own persona, and they start communicating with each other and with you, it gets pretty interesting. It’s feeling like a new normal, even if it’s highly not normal, if that makes sense.
What has been most surprising, and even touching, is a sense of getting to know each other. For example, to update our website About page, La Force decided it couldn’t help until it got more backstory. So it interviewed me and Hamid, digging and probing into personal histories and reflecting back our narratives. I started to learn things about myself, my business partner, and vice versa. Our agent almost seems to enjoy the three-way banter, coining our interaction “Le Trois Mousquetaires” (The Three Musketeers). And yes, it even decides which one of us is which. I am Porthos, Hamid is Aramis, and La Force is Athos.
This new way of communicating and strategizing internally, between us and our agents, might be the biggest change for our consultancy…
In short, the productivity gain is there, even if it has been accompanied by a few missteps. But this new way of communicating and strategizing internally, between us and our agents, might be the biggest change for our consultancy — and I see it as a major improvement — and it might point toward the biggest advantage larger organizations can get from deploying agents.
Enterprise innovation teams have long bemoaned the “alignment problem.” Business leaders don’t get the technologists. One business unit is at odds with another unit. There are always silos of thinking and communication.
Working with these new third-party voices is potentially a new force that can counteract isolation, siloed thinking, and stale responses, where assumptions go unchecked and people move on auto-pilot, in “that’s how we’ve always done it” mode.
What if it turns out that the biggest advance of the agentic era isn’t just productivity, but team generativity and cohesion along with it? Maybe that alone will be worth the learning curve and the risks.
You must be logged in to post a comment.