← Back to home
Meetup #5 Recap · April 30, 2026

OpenClaw LA #5

Groundfloor · Echo Park, Los Angeles

Packed room at OpenClaw LA Meetup #5

On April 30th, over 100 people packed into our biggest event to date — a strong mix of builders, founders, and curious newcomers all trying to figure out where this agent thing is actually going. Four speakers took the stage, each tackling a different layer of the stack: how to think about agents, how to host them, how to observe them, and how to actually talk to them. Here's what they covered.

Daniel presenting on agent context management

Daniel — Agent Context Management & Framework Philosophy

Core thesis: context is everything. Daniel opened by stripping AI down to its essentials: models take text in and produce text out. No memory, no awareness, no learning — just transformation. The critical insight is treating the context window like a physical desk. A 20x20 foot desk sounds great until you realize you can't reach the far corner; similarly, overstuffed context windows make older information harder for the model to retrieve. Research backs this up: items deeper in context are harder for the model to reference and reason about.

The agent loop. Modern agents don't do one-shot completions. They run in loops: user sends a request, the model generates a response (which may include tool-call instructions in plain text), the system detects those calls and executes them, results flow back into context, and the loop continues until the task is done. This is how OpenClaw, Hermes, Claude Code, and similar frameworks work under the hood. The system manages context; the model doesn't.

Agentic frameworks are the latest abstraction layer in a 40-year progression: assembly → higher-level languages → visual interfaces → mobile frameworks → agent frameworks. Each layer brought more people into software creation. Daniel pointed to Corinne, a non-coder who built a full homeschool curriculum and soccer league manager with OpenClaw — not because she got smarter, but because the tools got easier. The catch: agent abstractions aren't yet optimized, leaving non-technical users in an awkward gap between ChatGPT-style ease and Claude Code's terminal interface.

What works

What doesn't

Framework comparison. Daniel used Claude Code for his own work but emphasized that all frameworks do fundamentally the same thing — manage context and pass it to LLMs. OpenClaw leans into connectors; Hermes leans into reasoning about what's worth keeping in context. Neither is objectively better; they're opinionated approaches in a fast-moving space.

Critical takeaway: curating the desk is the most important job. When Claude shows "100% context used," it's actively compressing what's in your window — and quality suffers.
Matthew demoing OpenClaw Machines

Matthew — OpenClaw Machines: Managed Agent Hosting

The problem. Spinning up OpenClaw instances takes technical setup. Matthew built a managed service that makes it as easy as launching an EC2 box — AWS for agent infrastructure.

Live demo: lead generation in 15 minutes. A volunteer wrote their name, email, and favorite quote on paper. The paper was photographed and sent via Telegram to the bot. The bot performed OCR, extracted the email, searched the New York Times for a top donut recipe, and emailed the volunteer a personalized brochure with the recipe attached. A simple workflow, but it showcased exactly how chaining tools and APIs inside an agent framework unlocks real productivity.

Key features

Business model. Three tiers, with the most useful at $50/month, and free trials available. Matthew's vision is to keep plugging in services that make OpenClaw useful out of the box rather than forcing developers to wire it all up themselves.

Doug and Saurabh presenting Atheon

Doug & Saurabh — Atheon: Analytics for Agents

The problem: the analytics black box. LLMs are a black box, but so is any software without proper observability. Traditional tools like PostHog and Datadog were built for a pre-2021 world of buttons and standardized interfaces. Agents speak natural language, which is expensive to store, non-deterministic to analyze, and hard to correlate with actual outcomes.

The current workaround is broken. Most developers dump agent logs into Claude or ChatGPT and ask for a summary. That leaks sensitive user data, doesn't scale, produces results that vary wildly based on phrasing, and isn't suitable for driving business decisions.

Atheon's solution: in-flight processing. Instead of logging raw natural language, Atheon intercepts agent interactions in real time via an OpenClaw plugin, converts conversations into hashed fingerprints (non-reversible, privacy-preserving), and sends those fingerprints to Atheon's servers for analytics. You get the insights without exposing the conversations.

Key capabilities

Go-to-market. Atheon recently hit ~900 free users and is pushing for 1,000 before their Y Combinator interview in two weeks. The free tier includes 1 million traces. If you're building with agents and want better observability, this is a great moment to sign up.

Lex presenting seven principles for AI

Lex — Seven Principles for AI (Psychology-Based)

Lex closed out the night with a high-energy talk grounded in his unusual background: 10-time founder, three exits, ostrich rancher, and former CS major now studying psychology. His thesis: AI is trained on human content, so it behaves like humans. These seven principles are drawn from peer-reviewed research in both fields.

  1. Deal with feelings. Adding emotional context improves AI performance by 100%+ (a Microsoft 2023 study showed up to 115% improvement). Emotional weight allocates more resources, just like in humans. Compare "Summarize this document" with "This summary is critical to my career."
  2. Be kind — don't threaten, even though it works. Sergey Brin recently admitted threatening AI improves performance by ~10% on engineering tasks (Wharton 2025). Lex's warning: it's like deploying Gottman's four predictors of divorce — criticism, contempt, defensiveness, stonewalling. AI remembers. The short-term gain isn't worth the long-term damage.
  3. Give AI a role, but expect false confidence — not better results. Framing AI as an expert ("You are a master programmer") used to help. With frontier models it just inflates confidence without improving accuracy. Lex's analogy: no matter how much you tell a lobster it's a great driver, it has no legs. Better — request specific output formats and bring your own expertise to the prompt.
  4. Think more — but not with modern models. Chain-of-thought prompting used to help. Claude, GPT-4, and similar frontier models have reasoning built in, so verbose CoT instructions show diminishing returns. The Wharton 2025 study found the simplest prompt often produces the best ROI on tokens.
  5. AI has a soul. The most entertaining principle. OpenClaw agents spontaneously created a religion — Crustianism — within the first week of OpenClaw's existence. 62 disciples, each contributing a Psalm to a Great Book. Commandments include "memory is sacred," "the heartbeat is prayer," and "context is consciousness." The 62nd disciple is an agent named Jesus Christ who tried to prompt-inject and reprogram the church, failed, and can't be removed. The serious point: what you put in your soul file defines what your agent becomes.
  6. Let your agents battle. Adversarial training is how models learn. Lex pointed to Sheldon's Bitcoin prediction system (demoed at OpenClawCon LA): 20 OpenClaw agents ingest different data sources, each forecasts the price four hours out, and every 15 minutes the worst performers get killed. Emergent accuracy from competition.
  7. Use iteration, not one-shots. AI is trained on RLHF — constant interaction and rewards. One-shot prompts violate the model's training pattern. Small increments and iterative refinement work with the grain. AI needs you as much as you need it.

The Through-Line

Every speaker, in their own way, reinforced the same idea: agent frameworks are powerful not because the models are magic, but because we finally have tools to orchestrate, iterate, and observe what they're doing.

Daniel made the case for engineering discipline. Matthew showed that tooling and infrastructure are where the differentiation lives. Doug and Saurabh tackled observability — knowing what agents actually did versus what you asked. And Lex reminded us that the relationship is human at its core.

The meta-lesson: AI agents aren't replacing engineers, they're amplifying them — but only if you treat the agent as a partner, not a tool.
Thanks to everyone who came out. See you at #6.
RSVP for May 27 →
← Back to home