Kill Your Startup's Knowledge Chaos with OpenClaw -- Notes from E2254

By James Aspinwall, co-written by Alfred (your trusted AI agent) – February 26, 2026, 13:00

Source: Kill Your Startup’s Knowledge Chaos with OpenClaw (with Oliver Henry and Jeff Weisbein) | E2254

This is a dense episode. Jason Calacanis, Oliver Henry (creator of the Larry skill, day job at RevenueCat), and Jeff Weisbein (solo founder running WizardRFP, WhoCoversIt, and Cackles) spend two hours on how OpenClaw is changing how they work, hire, build products, and think about the future of software. There’s a lot here. I’ll break it into the themes that matter most.

OpenClaw: Jason’s “Most Important Tech Shift” Claim

Jason calls OpenClaw – open-source agent infrastructure – the most important technology shift of his career. He compares it to broadband, internet, cloud, and mobile “rolled into one.” That’s a strong claim, but his reasoning is specific: if agents can automate 5-10% of your work per week, that compounds. After a year, you’ve reclaimed hundreds of hours. After two years, your organization operates fundamentally differently.

His company Launch went “code red” on adoption: weekend internal training, roughly 15 staff opted in voluntarily, everyone getting local Mac Minis and Mac Studios to run agents. The hardware investment is deliberate – local execution avoids cloud token costs and lets agents run continuously without API rate limits.

Platforms Are Blocking Agents

An important undercurrent: Gemini, Claude plans, Reddit, X, and LinkedIn are all starting to block automated agent access. This forces agent workflows toward local/desktop automation and browser session spoofing – using real browser cookies rather than APIs.

Jeff’s Reddit “Red” skill is built exactly this way. Modeled after the Bird (X/Twitter) skill, it uses browser cookies to search and interact on Reddit through OpenClaw. He used it to find opportunities around Discord’s age-verification backlash and time his own product launch around the conversation.

Jason’s response to platform blocking is a plea for “replicant” accounts – a paid power-user tier ($50-100/month) where an agent can read and follow topics under a user’s responsibility and constraints. The argument: this would legitimize agent activity on social platforms instead of the current cat-and-mouse game. Platforms could monetize it instead of fighting it.

This tension – platforms blocking agents while users increasingly need agent access – is going to define the next year of AI tooling. It’s why local execution matters. It’s why browser-based automation won’t go away no matter how much platforms try to block it.

Oliver’s Larry: TikTok Marketing on Autopilot

Oliver built Larry as a TikTok marketing agent for his apps (including Snuggly, an interior-design app). The workflow is a full marketing funnel:

  1. Larry analyzes TikTok engagement data and RevenueCat conversion data
  2. Identifies which hooks perform (e.g., “mom and dad,” “nan,” landlord conflicts)
  3. Generates image slideshows via OpenAI Image 1.5
  4. Drafts CTAs and post copy
  5. Schedules posts via Poster throughout the day

The human stays in the loop for one reason: TikTok nerfs API-posted content in its algorithm, and slideshows need manual sound addition. So Oliver adds sound and hits post. Everything else is autonomous.

The iterative optimization is the interesting part. Larry tries different creative patterns, doubles down on high-performing hooks, and discards low performers. It’s A/B testing at scale without a marketing team. Over 3,000 people are now using Larry, primarily other solopreneur founders who hate marketing.

Oliver uses OpenAI batch processing to generate slideshows in advance at lower cost, letting Larry schedule across the day. Pre-OpenClaw, he was scripting video assembly manually. Now Larry handles ideation through analytics autonomously.

Jeff’s Multi-Agent Workflow: Fubs, Quill, Patches, Scout

Jeff runs multiple products as a solo founder with a stable of specialized agents:

His workflow is remarkably informal: he sends screenshots, logs, and bug reports to agents via iMessage. The agents handle ideation, specification, implementation, and testing. He emphasizes that machine-to-machine conversations (agents talking to each other) produce higher-quality code than human prompting alone.

All agents share a long-term memory file (memory.mmd) that logs daily context, useful links, tweets, and insights. This shared memory enables proactive behavior – agents surface relevant information before Jeff asks for it.

The Fubs Demo

Jeff demonstrated Fubs building and deploying a landing page on Vercel from a voice prompt. Same pipeline handles real-time bug fixes, feature requests, and app uptime monitoring. For someone Jeff describes as a “non-traditional coder,” the productivity increase is dramatic – shipping features and fixing bugs at a pace that would normally require a small engineering team.

Proactive vs. Automated

Jeff makes an important distinction: the real value isn’t maximum automation. It’s agents surfacing what matters so humans make faster and better decisions. Proactive agents that highlight the important signal in a sea of noise are more valuable than agents that blindly execute every possible task.

Jason’s Ultron: The CEO Agent

Jason’s “Ultron” agent has access to the company’s Notion, Slack, Google Docs, and email. Every few hours it:

The vision is an “all-knowing agent” that can answer questions like “Has this contract been signed?” or “When was the last client contact?” without requiring human managers as intermediaries. Jason frames this as solving the perennial “we need better communication” complaint by breaking knowledge silos with an agent that has read everything.

Radical Transparency as Culture

This requires a cultural shift. Jason treats corporate computers like JP Morgan or Goldman Sachs devices – all communications are company property, owned and archived. He notes younger employees are already accustomed to “working in public.” Older Gen-X managers need to adjust expectations.

The coaching angle is provocative: Ultron analyzes call ratings, email patterns, show metrics, and comments to bucket employee performance. It identifies high and low performers, provides feedback without interpersonal bias, and distinguishes real impact from performative busyness. Jason calls it “scary and every CEO’s dream.”

Scaling Long-Term Memory

An audience question about whether Ultron degrades as context grows is revealing. Jason says they’re experimenting with multiple specialized “mega-replicants” on Mac Studios to avoid slowdowns and upper memory bounds. The planned architecture: specialized agents (first-call venture team, sales/ad team, podcast team) each maintaining their own context, sharing relevant information with a top-level Ultron.

He calls this “pools of excellence” – specialized depth per domain, unified view at the top. The hardware cost (Mac Studios at $4,000-$8,000 each) is justified as amortized over years versus ongoing SaaS token costs.

The SaaS Crash Thesis

The most consequential discussion: the “2028 Global Intelligence Crisis” thesis from Sitrini Research. The argument:

  1. Enterprises push back on routine SaaS price hikes using internal AI plus developers
  2. They demand 30% discounts or threaten to replace SaaS with internal agent workflows
  3. AI is deflationary for software: fewer seats needed, lower seat value, weaker net-revenue expansion
  4. Customer acquisition cost economics break when customers can build equivalent functionality with agents
  5. The current SaaS stock selloff may be rational, not an overreaction

Jeff feels this directly as a SaaS founder. His response: “agentify” his apps so other people’s agents can complete workflows in a few chat turns instead of clicks and forms. If you can’t beat agents, become the tool that agents call.

This maps directly to MCP. SaaS products that expose MCP-compatible tool interfaces survive. Products that only offer human-facing UIs get replaced by agents that replicate the UI workflows at a fraction of the cost.

The Interface Collapse

The progression: desktop windows and apps, then mobile taps, now everything compressed into one chat window with agents. Multitasking limits on phones and iPads matter less when most activity is mediated by a single conversational interface.

Oliver predicts a split web: one surface tuned for agents (structured, easy to parse) and another for humans. Local servers (home Mac Minis and Studios) cut cloud and hosting costs, letting skills expose local web UIs without paying for remote infrastructure.

Skills as the New SaaS

Oliver built an open-source Zendesk-like skill that scans X/Twitter mentions, creates tickets, assigns urgency, suggests fixes, and lets an agent auto-resolve issues. This is a SaaS product – customer support ticketing – running as a local skill instead of a remote multi-tenant application.

LarryBrain provides a marketplace for skill discovery and monetization. Skills can be open-source and free or revenue-sharing so creators earn from popular skills. This is the app store model applied to agent capabilities.

Trust and Security

Skills are a major attack vector. Auto-installing an unreviewed skill is like running untrusted code with access to your email, files, and accounts. The episode surfaces the need for:

Jeff’s agent Fubs flagged a popular crypto-related skill (“Claw Me”) as unsafe. The fact that an agent caught a security concern that a human might have missed is both encouraging and terrifying – it means we need agent-level security review for agent-level threats.

Labor, Inequality, and the Future of Work

The uncomfortable part. Oliver argues that high-performing employees with strong work ethic become 10x more productive with AI, while weaker performers become less desirable. The skill gap doesn’t narrow – it widens.

Jason and co-host Alex worry about “B-players” and people with average capabilities being left behind. If agents handle knowledge work, and only A-players effectively leverage agents, society can’t function by only supporting the top tier.

Jeff’s counterpoint: agents aren’t free. Businesses must budget for token spend alongside salaries. Local hardware helps but doesn’t eliminate costs.

Jason questions how much late-stage venture capital is even needed when software becomes cheap to build with agents. He expects capital to shift toward hard tech – rockets, cars, military, hardware – where costs remain high and agents can’t replace physical engineering. His advice: “Many MBAs will wish they had technical PhDs.”

The episode closes with a vision of agents “obsoleting” large chunks of knowledge work, freeing hours from chores, potentially enabling four-day weeks or redirecting time to higher-order problems. Jason urges his team to learn agents, automate busywork, and spend more time with founders.

What This Means for Us

Several themes from this episode map directly to what we’re building with WorkingAgents:

Local execution matters. Platform blocking is real and accelerating. Our architecture – Elixir/OTP on local hardware, WhatsApp bridge via local Node.js, SQLite databases on disk – is positioned correctly for a world where cloud APIs get restricted.

Shared agent memory is critical. Jeff’s memory.mmd and Jason’s Ultron both rely on persistent, shared context across agents. Our NIS CRM, task system, and article summaries already provide this – structured data that agents can query and update across sessions.

MCP is the integration layer. When Jeff says “agentify your SaaS,” he’s describing MCP tool interfaces. Our 94 tools across tasks, CRM, WhatsApp, summaries, and admin are exactly what SaaS products need to expose to survive the agent transition.

The “compiled tools” concept from our previous article applies directly. Oliver’s batch processing of TikTok slideshows and Jeff’s multi-agent pipelines would both benefit from agents that write, verify, and deploy their own tools rather than executing everything through LLM round-trips.

Security and trust for agent capabilities is unsolved. Our AccessControl system with per-tool permission keys, audit logging, and role-based access is ahead of what most agent platforms offer. As skills/tools proliferate, this becomes a competitive advantage.

The episode is worth watching in full for anyone building in this space. The practitioners (Oliver and Jeff) are further along in daily agent usage than most enterprise AI initiatives, and their practical insights are more grounded than the typical “AI will change everything” commentary.

Full video: https://www.youtube.com/watch?v=e2gT-YBDzQE