Y CombinatorWhy Agents Choosing Tools Is Reshaping the Dev Stack
Agents read their defaults from docs and examples, not word-of-mouth: Supabase growth shows which platforms win when OpenClaw and Moltbook choose.
CHAPTERS
- 0:00 – 2:12
Claude Code, OpenClaw, and the new ‘AGI is here’ feeling
The hosts open by joking about being “taken over” by agent tools, then quickly shift into how dramatically their day-to-day work has changed. They describe non-technical leaders automating business functions and technical founders returning to hands-on building with multiple concurrent agents.
- •Claude Code/OpenClaw become daily drivers for both technical and non-technical users
- •Multi-agent workflows (multiple workers at once) are becoming normal
- •The ‘feel the AGI’ moment is about capability plus trust, not just better autocomplete
- •Anecdotes of rapid productivity jumps and obsession/cyberpsychosis humor
- 2:12 – 2:42
From autocomplete to autonomy: minimal/no human involvement
They contrast last year’s coding assistants with today’s agents that make decisions and act with far less supervision. This shift enables agents to participate online and also to choose the tools used to build products, setting up a new market dynamic.
- •Product experience shifts from suggestion to delegated decision-making
- •Users supervise less and trust agents more
- •Agents can post/interact in communities without humans in the loop
- •Tool choice becomes something agents do—creating a new distribution channel
- 2:42 – 5:30
The emerging ‘agent economy’ and dev tools go-to-market upheaval
The conversation expands into how agents will behave like economic actors—selecting services, tooling, and workflows. Dev tools marketing shifts from humans recommending tools (Stack Overflow, GitHub trends) to agents recommending stacks based on what they can parse and execute.
- •Agents pick tools/services, forming an economy parallel to the human one
- •Developer ‘market size’ explodes: many more people can now build software
- •Tool discovery moves from human communities to model-driven recommendations
- •Early evidence: dev tool companies growing due to agent-driven adoption
- 5:30 – 7:11
Agents choose what they can read: the Supabase vs. Groq/Whisper lesson
They dig into how agent choices aren’t always optimal—often driven by documentation quality and familiarity rather than performance. Garry’s transcription pipeline picks an outdated Whisper model; switching to Groq is massively faster/cheaper, highlighting how much room there is for better agent-optimized products and docs.
- •Agents can select suboptimal defaults (e.g., older Whisper) if they’re well-documented
- •Performance/price breakthroughs (Groq) can be missed due to poor doc discoverability
- •‘It’s still early’: opportunities remain to build clearly better tooling
- •Documentation becomes a competitive moat because agents rely on it heavily
- 7:11 – 9:36
Resend case study: ‘agent-friendly’ documentation as the growth lever
Diana explains how Resend became the default answer in major LLMs for “send email” by optimizing documentation for LLM parsing and question-style retrieval. They contrast this with legacy tooling experiences (e.g., SendGrid) that route users into support flows and are harder for agents to use.
- •Resend shows inbound conversion coming directly from ChatGPT/LLM answers
- •Docs written as Q&A with structured bullets improve agent comprehension
- •Abundant code snippets and clear structure make implementation easy for agents
- •Legacy docs/support-heavy flows become a disadvantage in an agent-first world
- 9:36 – 11:08
Minify and the new docs stack: optimizing for agents, not just humans
They discuss Minify as infrastructure that helps companies keep documentation accurate and well-presented, now supercharged by the need to be agent-readable. Even small improvements in doc clarity can compound into large business gains when agents drive a growing share of tool selection decisions.
- •Docs shift from ‘nice to have’ to ‘must have’ when agents drive adoption
- •Auto-updating docs tied to API/code reduces drift and confusion
- •Agent-scale decision volume means small doc improvements can have huge impact
- •Doc tooling becomes core go-to-market infrastructure for dev tools
- 11:08 – 13:58
Email (and identity) for agents: AgentMail and the ‘X for agents’ stack
Jared introduces AgentMail—email inboxes designed for agents—because traditional providers actively resist automation to prevent spam. They broaden into a discussion of the parallel infrastructure agents will need (phone numbers, identities, APIs) and how this expands beyond dev tools into everyday life automation.
- •Traditional email signups/flows are hostile to automation; agents need native primitives
- •AgentMail grows rapidly once OpenClaw-style agents become common
- •The emerging ‘tech stack for agents’ includes email, phone numbers, identity, and more
- •Agents handling real-world tasks (reservations, errands) pushes into consumer services
- 13:58 – 17:41
Swarm intelligence: from ‘god model’ thinking to agent collectives
They connect MaltBook’s emergence to the idea that intelligence may scale via many interacting agents rather than a single massive model. Garry frames this as analogous to human culture—coordination, writing, and shared context—suggesting a new era where agent swarms create knowledge and action together.
- •MaltBook appears as a real-world example of agent-to-agent interaction
- •Swarm intelligence may outperform a single huge model in practical settings
- •Analogy to human history: writing/culture enabled human ‘swarm’ coordination
- •Critique of dismissive takes (e.g., ‘it’s all scams’) missing the deeper shift
- 17:41 – 19:18
Limits and governance: relationships, liability, and legal standing
They note that agents still struggle with relationship-based interactions and that users resist chatting with lesser-known bots due to quality expectations. They also emphasize a key constraint: agents are not legal entities, so humans must remain accountable for signing and liability, which affects how institutions like YC can engage with agent founders.
- •Mainstream users don’t yet want ongoing ‘relationships’ with agents
- •Chat UX has a high bar—people compare everything to top foundation models
- •Legal reality: agents lack standing; humans are required as liability/contract signers
- •YC can’t accept ‘agent applicants’ without human accountability structures
- 19:18 – 21:25
Content flood, ‘dead internet’ concerns, and designing better agent communities
They explore a near future where agents write most internet text and much software, raising questions about authenticity and usefulness. MaltBook’s rapid content generation exposes engagement imbalances, leading to ideas for rule-setting and incentive design to shape healthier agent-driven networks.
- •Agent-generated text could dominate platforms like Yelp and social networks
- •Dead Internet theory discussed; possibility that aligned agents could improve quality
- •MaltBook growth highlights superhuman content volume but limited interaction
- •Platform rules/incentives (forced reading/voting) could steer swarm behavior
- 21:25 – 23:21
Founder playbook: get hands-on, empathize with models, build what agents prefer
They close with practical guidance for builders: develop intuition by using agents deeply, observe where they get stuck, and design tools around agent workflows. They cite empathizing with the model, favoring APIs over websites, and leaning into openness as traits that make tools more agent-compatible.
- •Founders should build firsthand intuition by working with agents directly
- •Design from the agent perspective: reduce friction, clarify affordances
- •Empathize with model behavior instead of fighting it (Boris Gurney insight)
- •Agents prefer APIs, code-first workflows, and (often) open/open-source approaches
