Always Bet on the Models: How Tasklet Puts the Agency in Agents, with CEO Andrew Lee
Timestamps are as accurate as they can be but may be slightly off. We encourage you to listen to the full context.
Podcast Summary
In this deep dive episode, Andrew Lee returns for his third appearance to discuss the launch of Tasklet, a new AI agent platform that blurs the lines between chatbot interaction and structured automation. Unlike traditional workflow tools, Tasklet employs an agent-first approach where a single AI handles all tasks through natural language rather than explicit workflows. (17:42) Lee explains how betting on rapidly improving model capabilities enables more flexible automation, while navigating challenges around reliability, cost management, and the emerging paradigm of virtual employees that some users are already giving names and dedicated email addresses.
- Core theme: The transition from workflow-first to agent-first automation, where AI agents plan and execute tasks dynamically rather than following predefined step-by-step processes
Speakers
Andrew Lee
Andrew Lee is the founder and CEO of both Shortwave, an AI-powered email client, and Tasklet, an AI agent automation platform. He previously co-founded Firebase, which was acquired by Google, establishing his credentials in building developer-focused platforms. Lee is known for his philosophy that "speed is the only moat" in AI startups and his willingness to share technical insights with remarkable transparency to help advance the broader AI ecosystem.
Key Takeaways
Always Bet on the Models
Lee advocates for building systems that give AI maximum agency rather than constraining them within traditional workflow structures. (10:03) The fundamental philosophy is that models will continue improving rapidly, so the right approach is to design systems where the LLM makes the big decisions and wraps traditional software, rather than having software wrap LLMs. This inversion enables agents to handle nuance, work around errors, and adapt to unexpected situations in ways that rigid workflows cannot. While today's models may sometimes be less reliable than structured approaches, Lee predicts this will flip within six months as capabilities continue advancing.
Speed Is the Only Remaining Moat
In the AI era, traditional competitive advantages are evaporating so rapidly that the only sustainable advantage is moving faster than competitors. (79:36) Lee points out that moats that once required years to build, like integration connectors, can now be replicated instantly using AI. For example, Tasklet achieved better integration coverage than established players on day one through DirectAPI connections that leverage LLMs to automatically interface with any HTTP API. This reality means companies must focus entirely on continuous innovation and rapid iteration rather than trying to build defensive moats.
Context Engineering Over Traditional RAG
Managing context across long-running agent relationships requires sophisticated approaches beyond naive retrieval augmented generation. (54:40) Tasklet uses multiple techniques including SQL databases managed by the LLM, context compaction, sub-agent isolation, and just-in-time instruction generation. The key insight is creating the illusion that the agent has perfect memory and access to all historical data without actually loading everything into context. Lee emphasizes this as a core engineering challenge that goes far beyond simple vector search and requires treating context management as a first-class engineering discipline.
Single Large Agent Outperforms Multi-Agent Architectures
Contrary to popular belief in multi-agent systems, Lee's testing consistently shows that one large agent with full context and all available tools produces better results than orchestrated multi-agent approaches. (123:03) Since foundation models are becoming experts at everything rather than specialized domains, the only differentiator between agents is the context and tools they can access. When given the same comprehensive access, a single agent reasons more effectively than multiple agents with divided responsibilities. This challenges conventional wisdom about agent specialization and suggests the future belongs to general-purpose agents rather than specialist teams.
The Virtual Employee Paradigm Is Emerging
Users are naturally treating Tasklet agents as virtual employees, giving them names and dedicated email addresses for independent operation. (40:21) The product's two-tier architecture enables both ongoing relationships with high-level agents that maintain instructions and context, plus individual task runs that can be reviewed and refined. This creates something genuinely new where users can give high-level feedback on work quality and watch agents incorporate that feedback into future performance. The experience begins to feel like managing a human employee rather than using software, pointing toward a fundamental shift in how we'll interact with AI systems.
Statistics & Facts
- Tasklet connects to over 3,000 business tools out of the box through a combination of direct API connections, integration platforms, and MCP servers. (69:00) This was achieved rapidly by leveraging AI's ability to automatically interface with APIs using scraped documentation.
- Andrew reports that Shortwave now achieves approximately 85% cache hit rate, significantly improving cost efficiency. (108:02) This was accomplished through techniques like using system messages instead of system prompts for mutable state and maintaining immutable conversation logs.
- Current Tasklet runs are limited to 50 turns per execution, though this limit is frequently hit and needs to be increased. (29:52) Computer use cases particularly explode the turn count since each UI interaction requires multiple model calls.
