How to code with AI agents - Advice from OpenClaw creator | Peter Steinberger and Lex Fridman
TL;DR
Steinberger details his evolution to an 'agentic engineering' workflow using multiple CLI-based AI agents simultaneously, arguing that mastery requires developing empathy for how agents perceive limited context while embracing imperfection and concise prompts over complex orchestration.
🖥️ The Shift to Agent-Driven Development 3 insights
Terminal-First Workflow
Transitioned from IDE-heavy work to running 7+ Cloud Code terminal windows side-by-side, using the IDE only as a diff viewer and rarely reading code directly.
Selective Code Review
Stops reading 'boring' boilerplate (data transformation, Tailwind alignment) and focuses human review only on critical architecture, database logic, and security-sensitive PRs.
Voice-Native Input
Uses voice-to-text extensively for agent prompting (reserving hands for terminal commands), having temporarily lost his voice from overuse, finding spoken language more natural than typing.
🧠 The Agentic Trap and Engineering Philosophy 3 insights
The Complexity Curve
Engineers progress from simple prompts to over-engineered multi-agent orchestration with custom workflows, then reach 'zen' by returning to short, simple prompts like 'look at these files and make these changes.'
Agentic Engineering vs Vibe Coding
Rejects 'vibe coding' as a slur implying carelessness; 'agentic engineering' requires practiced skill, treating AI like a capable junior developer who sometimes needs guidance but often has better ideas than the human.
Empathy for the Agent
Success requires understanding that agents start each session with zero context and limited context windows, necessitating guidance on where to look rather than assuming full codebase knowledge.
⚡ Modern Workflow Practices 3 insights
Always Commit Forward
Never reverts code; instead fixes issues forward with agents, maintaining a 'YOLO' approach where main is always shippable and refactors are cheap enough to do on demand.
Conversational Debugging
Treats interactions as discussions with agents, asking 'do you understand the intent?' before implementation, and stopping long-running tasks to reassess architectural friction rather than forcing solutions.
Local-First Validation
Runs CI locally before pushing rather than relying primarily on GitHub CI, prioritizing speed and iteration over traditional branch protection models.
🤝 Human-AI Collaboration Balance 3 insights
Letting Go of Perfection
Accepts that agents produce 'good enough' code that differs from personal style, comparing it to managing human engineers where micromanagement destroys velocity and morale.
Codebase Design for Agents
Structures projects using obvious, searchable naming conventions that agents can discover, rather than optimizing for human aesthetic preferences that confuse AI search patterns.
Against Full Automation
Opposes orchestrators like GitTown that attempt to automate the entire loop, comparing them to failed waterfall models; believes human vision and style require iterative, hands-on involvement.
Bottom Line
Treat AI agents like capable junior engineers with fresh perspective—provide short, clear guidance that accounts for their limited context window, accept working but imperfect solutions rather than forcing your worldview, and maintain human control over architectural vision instead of attempting full automation.
More from This Week in Startups (Jason Calacanis)
View all
The limits of AI scaling laws - NVIDIA CEO explains | Jensen Huang and Lex Fridman
Jensen Huang explains that AI progress is now driven by four simultaneous scaling laws (pre-training, post-training, test-time, and agentic), with synthetic data eliminating previous data scarcity concerns and making compute the sole limiting factor for intelligence growth.
Origin story of OpenClaw: From 1-hour prototype to 180,000 stars of GitHub | Peter Steinberger
Peter Steinberger explains how a 1-hour WhatsApp-to-CLI prototype evolved into OpenClaw, the fastest-growing GitHub repository in history (175,000+ stars), by creating a self-modifying AI agent that prioritizes fun and accessibility over corporate polish.
The "secret sauce" of recent AI breakthroughs: Post-training with RLVR (and RLHF) | Lex Fridman
Recent AI breakthroughs in reasoning models stem from Reinforcement Learning with Verifiable Rewards (RLVR), which trains models by rewarding accurate solutions to objectively checkable problems like math and coding, enabling scalable performance gains through iterative trial-and-error rather than human preference optimization.
Timeline to AGI: When will superhuman AI be created? | Lex Fridman Podcast
The conversation contrasts the "AI 2027" report's milestone-based path to AGI (superhuman coder → researcher → ASI by 2031) with the "jagged capabilities" view, concluding that while AI will automate significant software development tasks within months, fully autonomous research and general computer use remain distant due to specification challenges and uneven capability profiles.