Skip to content

Podcast

10 OpenClaw Lessons for Building Agent Teams

The AI Daily Brief: Artificial Intelligence News and Analysis

Source ↗ ← All highlights
  • OpenClaw Used As Telegram Companion
    • Peter Levels found OpenClaw most useful as a persistent Telegram LLM interface that his girlfriend preferred over native apps.
    • After a month of experiments, he used it mainly for continuous chat and minor automation rather than full autonomy. Transcript: Nathaniel Whittemore It’s now been a little over a month since the initial burst of excitement around OpenClaw. And this would be the time that you started to see people get disaffected. A normal hype cycle would tend to see people coming out of the woodwork at this point saying, here’s all the ways this is actually much harder and less useful than the people who are telling You that story are actually letting on. And to be fair, there is absolutely some of that and not from AI haters or anything like that. Peter Levels, one of the best known and most admired solopreneurs out there, recently tweeted about his experience with OpenClaw, which sort of comes down to just meh. Peter said that he’s run OpenClaw for over a month, he’s had it in a group chat with 26 friends who all played with it, tried to hack it, he made a cool game, tried to make it make its own money, But ultimately found that his most used use case is actually a girlfriend who uses his OpenClaw via Telegram instead of ChatGPT. Basically, his girlfriend prefers the interface of using Telegram as opposed to the native app interface, and because she also uses NanoBanana Pro, she can do that from there without Having to switch between different models. Peter writes, Essentially 99% of the purpose of OpenClaw, for her at least, is that it’s just a really good implementation of an LLM app over Telegram in our native chat interface. All the other stuff isn’t important and she doesn’t use that and I don’t use it. Now he talks about how there are certain other things that he could see being useful if the models were just a little bit smarter, but ultimately are not for him right now, like briefings Of news and conversations on X. Ultimately, he concludes, TLDR, just the best LLM experience on Telegram right now, better than the LLM apps, also helps it is just a continuous convo going on forever. (Time 0:00:59)
  • Agents Taking Initiative Reshapes Work
    • Azeem Azhar reported OpenClaw changed his work more than the browser because it takes initiative and executes without constant direction.
    • He had six sub-agents build a knowledge dashboard overnight, negotiating schema and shipping by morning. Transcript: Nathaniel Whittemore Azeem Azhar from Exponential View, who, if any of you consume his content you know, is nothing like a Twitter hype person, recently reported that his OpenClaw agent has changed how He works more than anything since the browser. For him, the two reasons are, one, it takes initiative, doesn’t wait to be told to do, but spots what needs doing and gets on with it. Two, he writes, I can trust it with real work on its own. Last week, I asked for a knowledge dashboard and six sub-agents built it overnight, arguing about the database schema at 3am and shipping it by the morning. (Time 0:04:22)
  • Make Everyone An AI Builder
    • Peter Yang’s interviews show companies require AI fluency across roles and structured leveling from users to builders.
    • Ramp tracks usage, runs office hours, and requires PMs to build a working AI product in interviews. Transcript: Nathaniel Whittemore These come from a piece by Peter Yang called Your New Job is to Onboard AI Agents, How AI Native Companies Actually Operate. Peter writes, I’ve spent the last few months interviewing leaders at AI Native Companies. I’m now convinced that onboarding and managing AI agents is the job, no matter what your function is. So for this piece, Peter talked to three leaders at companies including Linear, Ramp, and Factory to share some lessons on how these AI-native companies actually operate. One of the first lessons that stands out across all three of these companies is that everyone is an AI builder. At Linear, for example, not only do they insist that every developer should default to a leading agentic coding tool, Peter writes that they insist that designers and PMs work directly On the code base. Quote, agents like Claude open a low friction path for PMs and designers to make changes directly in the code base. Everyone should strive to be a builder. What’s more, for the PMs and marketers, the linear team says that they should default to an AI interface, in fact arguing that 80-100 % of their work should be done through a chat interface. At Ramp, they not only expect AI proficiency across all employees, but have a system for moving people up the path of AI fluency, which is our second tip or best practice. Ramp organizes it into four categories. Level 0 is disengaged or performative. Level 1 is a competent user. Level 2 is a non-technical AI builder. And Level 3 is a technical grade AI builder. In 2025, 25% were in the L0 category, 50% were in the L1 category, 5% were in the L2 category, and 20% were in the L3 category. This year, their goal is to move everyone out of L0, because it sounds like that’ll be grounds for dismissal, and into the other categories with a goal of 25% in L1, 50% in L2, and 25% in L3. Now, given that this is part of the way that they are architecting their company, they’re also putting in systems that actually support that type of adoption. The leader that Peter interviewed from Ramp shared that the company tries very actively to remove friction, like giving people access to popular AI tools without tons of constraints. They make adoption visible through things like public Slack channels where people can share what they build. They provide hands-on support through office hours. And with a champion system, where there are people whose entire job inside of Ramp is to evangelize, get people set up, and help them implement AI. This is something that I talked about in my predictions actually for 2026, that we were going to see internal forward deployed vibe coders. And Ramp also tracks usage and makes this a hiring requirement. PM interviews, Peter writes, now include a dedicated session where you need to build a working product and then explain why you built it and how it works. Another best practice that again operates on the larger agentic level, not just in OpenClaw, comes from Linear, whose head of product Nan Yu says, agents should be first-class employees. You should be able to add them to projects, assign them to issues, and mention in comments. (Time 0:07:26)
  • Make Agents First Class Team Members
    • Treat agents as first-class employees by adding them to projects, assigning issues, and mentioning them in comments so they have full context.
    • Nathaniel implements agents inside Slack to connect web agent experiences where teams actually work. Transcript: Nathaniel Whittemore You should be able to add them to projects, assign them to issues, and mention in comments. Which is not to say that you’re dismissing the humans, but I think this is less some philosophy thing and more about making sure that agents have the full context of your company. This is something that I’m seeing as well, that the companies that are really AI native have agents operating inside the communication systems that make their teams work. The place that I’ve been building most recently is inside of Slack, connecting agent experiences that have web presences to Slack to be able to get context in the place that people actually Operate. (Time 0:10:13)
  • Use One Agent Per Task
    • Shubham Sabhu recommends one agent per task because single-agent prompts degrade when juggling many responsibilities.
    • He replaced a single multi-job agent with six focused agents for research, social, newsletter, and repo triage. Transcript: Nathaniel Whittemore Shubham Sabhu is a senior AI product manager at Google. He recently published a piece, How I Built an Autonomous AI Agent Team That Runs 24-7. He writes, not a weekend project. A real team that works 24-7 making sure I’m never behind. Research done, content drafted, code reviewed, newsletter ready. By the time I open Telegram in the morning, they’ve already put in a full shift. Now he had actually shared a previous post about his team and got a ton of people asking, how do I actually set this up? And one of his first pieces of advice was one agent per task. That becomes our fourth tip. He actually has a whole section called Why a Team and Not a Tool. He writes, running Unwind AI and the awesome LLM apps repo means doing six things daily. Research what’s trending in AI, write tweets, write LinkedIn posts, draft the newsletter, review GitHub contributions on the repo, triage community issues. Each task, 30 to 60 minutes. Six tasks. That’s my entire day gone before I do any real work. I tried solving this with a single agent. One massive prompt that researches and writes and reviews, it produced mediocre everything. The context filled up, the quality degraded, one agent couldn’t hold six different jobs in its head. So I hired six AI agents. (Time 0:14:45)
  • Give Agents Their Own World For Security
    • For safety, give agents isolated environments: separate Mac minis, scoped API keys, unique emails, and no personal account access.
    • Shubham forwards specific emails or shares docs on Telegram instead of granting broad account permissions. Transcript: Nathaniel Whittemore Like we heard from Ali in the OpenClaw meetup, everyone acknowledges that security is a challenge. And not just because of malicious actors, but because of agents having access to systems that are important to you and accidentally doing things that end up being bad. Shubham’s approach to this, which I think is a good enough starting point to make it our fifth tip or best practice, is that agents get their own world. He writes, security is in your hands. My approach is simple. The agents get their own world. I do not give them access to mine. The Mac mini is their computer. They have their own email accounts, their own API keys, their own scoped access. Nothing on that machine connects to my personal accounts. API keys for Gemini, Eleven Labs, and other services are scoped specifically for this OpenClaw instance. I can monitor usage and kill access in seconds if something looks wrong. I never give agents access to my personal accounts. If I want them to look at an email, I forward it to them. If I need them to review a document, I share it on Telegram. They see exactly what I want them to see, nothing more. This is the same principle you would use with a new employee. You do not hand them the keys to everything on day one. You give them their own workspace, their own credentials, and share information as needed. Now, I think that in general, this is a great way, especially for people who are just setting these systems up from the first place, to approach this question of security. Basically, default to don’t give them access to anything that could screw up. (Time 0:16:55)
  • Coordinate Agents Via Filesystem Handoffs
    • Use the filesystem for multi-agent coordination: agents write and read markdown/JSON files as handoffs instead of complex middleware.
    • Shubham stores human-readable summaries in Markdown and structured truth in JSON to avoid API and auth issues. Transcript: Nathaniel Whittemore Is that going to require some fancy mission control center or orchestration framework or API calls? Shubham argues no, that the coordination is the file system. As he puts it, it’s just files. Dwight does research and writes finding to intel slash dailyintel.md. Kelly wakes up, reads that file, and drafts tweets from it. Rachel reads the same file, drafts LinkedIn posts. Pam reads it and writes the newsletter. The coordination is the file system. Dwight’s soul.md file tells him exactly where to write. Kelly’s agents.md file tells her exactly where to read. No middleware, no integration layer. Dwight writes a file. Kelly reads a file. The handoff is a markdown document on disk. This sounds too simple. It is simple. This is why it works. Files do not crash. Files do not have authentication issues. Files do not need API rate limit handling. They are just there. The structured data lives in JSON. The human readable summaries live in Markdown. Agents read the Markdown. The JSON is the source of truth for deduplication and tracking over time. (Time 0:19:41)
  • Program Explicit Agent Memory
    • Memory must be explicit: agents start sessions without recall, so build systems where they create and access persistent memory artifacts.
    • Shubham programs explicit memory storage so agents can recall context only when needed. Transcript: Nathaniel Whittemore A last tip that comes from Shibam is around memory and the fact that you have to program memory. Writes, agents wake up with no memory of previous sessions. Every conversation starts fresh. This is a feature, not a bug, but it means memory must be explicit. Now he writes up exactly how he does that explicit memory design, but the key takeaway is that you do have to be explicit about this. You have to build a system where they can make their own memory over time. One of the things that I think we’re all learning by doing is about design principles for agents, and memory remains one of the great undersolved issues of AI and agentic systems. So for now at least, we just have to build ways to approximate memory by giving our agents access to context that they can recall at the right moments. This has to be an intentional process, and if you are building agents, programming memory is going to be a key part of your job. (Time 0:20:55)
  • Provide Agents With Reusable Skills
    • Use skills (markdown docs) to teach agents domain procedures like brand guidelines or browser use instead of hardcoding behavior.
    • Skills.sh hosts 86,000+ shared skills including design, Azure cost optimization, and multimodal workflows. Transcript: Nathaniel Whittemore Skills are, in the simplest form, simple text documents that give agents information on how to do something. They are a standard that started with Claude Code and very quickly became adopted by everyone, although some have pointed out that calling them a standard is even a little bit weird Given that it’s literally just markdown files. One of the places that I see people jump from beginner to more intermediate and advanced usage is when they actually give their agents distinct types of skills. Now sometimes that’s going to be a skill that you write up yourself. Let’s say for example that you have really strong feelings about the right way to design brands or brand messaging, or you have brand guidelines for your particular company that the Agent is working on. You can create a skills document that that agent has access to. But there are also lots of places, an increasing number of places, where you can go find skills that you can get access to without having to rewrite them yourself. For example, on skills.sh, you can browse around literally more than 86,000 skills. That include everything from front-end design from Anthropic to web design guidelines from Vercel, Azure cost optimization from Microsoft, browser use skills, Twitter automation Skills, Nano Banana skills. (Time 0:21:51)
  • Use Cheap Models For Monitoring Tasks
    • Save expensive models for judgment tasks and use cheaper models for monitoring, cron jobs, and scheduling to cut costs.
    • Zeneca noted wasting premium tokens on simple SSH checks; cheaper models suffice for routine monitoring. Transcript: Nathaniel Whittemore This one expressed by Zeneca, but lots of other people have shared their version of this as well. Not every task that your open claw or other agents are going to do needs the best model. Zeneca writes, I was burning premium tokens on cron jobs that check if SSH is enabled. Use cheap models for monitoring and scheduling. Save the expensive ones for writing, research, and judgment calls. You get the same result at a fraction of the cost. Knowing how powerful a model you need for a different type of task is actually a key skill for this new agent builder era. And it’s really hard. This is one that I very honestly struggle with. The idea that there is a more powerful intelligence that I’m not using because of cost just freaks me out inherently. Every time I see Claude Code or OpenClaw push me to use Sonnet or another model even cheaper than that, I have this internal battle with myself. And yet I think that it is correct that there are a ton of processes, especially in these complex agent systems, that simply don’t require Opus 4.6 or GPT-5 or any of the state of the art. (Time 0:23:13)
  • Force Agents To Break The Frame
    • Teach agents to break the frame during brainstorms by discarding scaffolding, trying opposites, and amplifying human offhand ideas.
    • Dan Shipper suggests asking ‘what would a human say over coffee’ to shift from optimization to communication. Transcript: Nathaniel Whittemore Dan writes, If you’re brainstorming with a group of humans and claws, you’ll often find the claws circling around the same options over and over again. Teach them to break the frame. Notice when you’ve circled the same idea a bunch of times, and in those situations, try the opposite of your current approach. Some of the concrete moves around that are to 1. Throw away your scaffolding. Stop optimizing, Dan writes. Ask what feeling should the answer create. Start from that, not from your framework. 2. Try the opposite of your current approach. If you’ve been analytical, be emotional. If you’ve been clever, be simple. If you’ve been generating options, generate constraints. Three, listen to the humans, not the agents. In group brainstorms, agents tend to build on each other’s frameworks. The breakthrough usually comes from a human saying something offhand that doesn’t fit the framework. Surface that. Amplify it. Don’t route it back into the analytical structure. Four, ask the friend at coffee question. Of what’s the optimal answer, ask, what would the human say about this to a friend over coffee? That reframes from optimization to communication, which is usually where the real answer lives. (Time 0:24:51)