Shortwave Rides the Tidal Wave: Inbox Agents, Hyper-Growth & Hiring AI Managers, with CEO Andrew Lee

Episode Summary

Executive Summary: Andrew Lee says Shortwave has evolved from a useful AI email assistant into an agentic communication platform that can search, triage, draft, and act across inboxes with surprising reliability. He credits rapid model progress, iterative tool use, Anthropic caching, and simplified hybrid search for better performance and margins, and argues the AI era rewards small, fast teams managing agents rather than traditional software orgs.

Main Topics: Shortwave’s product transformation from assistant to agent (Priority: 5/5): Shortwave moved from basic email Q&A and drafting to a more capable agent that can organize inboxes, compile reports, summarize threads, manage to-dos, and handle multi-step communication workflows with much higher trust and utility. Technical rebuild: search, retrieval, and model stack (Priority: 5/5): Andrew says essentially every layer was rewritten: embeddings, vector DB, search APIs, agent code, and model choices. The current stack uses Pinecone serverless, BGE embeddings, hybrid keyword-plus-semantic search, and more iterative model reasoning. Agent architecture and iteration over multi-agent complexity (Priority: 5/5): Shortwave found that a simpler single-agent approach with strong tools, repeated model calls, and caching outperformed brittle multi-agent or pipeline-based systems. Iteration and feedback loops are the core of effective agents. Costs, caching, and vendor strategy (Priority: 4/5): Anthropic’s caching is central to economics for long-context agent loops, reportedly cutting costs by around 90% after caching. Shortwave also uses different models for different tasks, but avoids over-engineered model chaining when it harms reasoning. Trust, safety, and next-step automation (Priority: 4/5): The company is expanding toward more autonomous actions like AI filters and eventually agent-driven actions on emails, but it must solve prompt injection, guardrails, and user trust before letting AI act more broadly without approval. Company structure for the AI era (Priority: 5/5): Shortwave is reorganizing around a small, high-density, in-person team that manages AI agents rather than doing all execution manually. Hiring emphasizes AI fluency, creativity, speed, and problem framing over traditional IC output. Broader software-industry implications (Priority: 4/5): Andrew argues AI is eroding the old moat of software code and making speed the main moat. He expects much more software to be built, but by smaller teams, with a shift in developer value toward product understanding and orchestration.

Key Arguments: Model progress over the last year fundamentally changed what agents can do; older attempts at tool use and multi-step reasoning were far less reliable. The best agent architecture for Shortwave is not many specialized sub-agents, but one strong model repeatedly called with good tools, good prompts, and iterative feedback. Hybrid search combining constrained full-text search with semantic vector search is cheaper, faster, and more accurate than older brittle custom retrieval pipelines. Anthropic caching is economically decisive for long-context, multi-step agents; without it, the product would be unprofitable at scale. The most useful AI experiences are those that turn messy communication history into concrete actions: triage, summaries, receipts, inventory reports, and custom-written replies. Shortwave’s moat is no longer the email client itself but execution speed, product iteration velocity, and the ability to build around rapidly improving models. Hiring should prioritize people who can manage AI well, frame problems, and think productively with tools, rather than only those who can execute tasks directly. AI should augment and sometimes automate routine work, while humans supervise high-stakes actions and handle nuanced judgment until trust and guardrails improve further. The future of communication software is omni-channel: email, Slack, LinkedIn, CRM, and other inbox-like surfaces should all be reachable through one AI layer.

Data Points: Revenue growth curve: Described as “beautifully exponential” / “going totally vertical” - Used to illustrate strong word-of-mouth-driven product-market fit Team size target: About 15 employees - Shortwave plans to stay very small for the foreseeable future Referral bonus: $10,000 - Offered for referrals to help Shortwave hire Cost reduction from Anthropic caching: Up to 90% cheaper - Caching long repeated agent contexts dramatically lowers model cost Agent iterations per task: Up to 20 tool/model calls - Illustrates how Shortwave’s agent loops through searches and actions Support thread example: 19 users - AI reported how many people complained about a new UI layout in 24 hours Retention on AI filter feature: 99% retention - Opt-in experiment used to judge whether the feature was working well Plan context window / data scope: Largest plan indexes all history and gives biggest context window - Customers willingly pay more for better answers and broader memory Model vendor usage: 3 vendors - Shortwave uses models from Anthropic, OpenAI, and Meta/Vertex depending on task Tooling cost expectation: Potentially 100x compute for full-agent-per-email automation - Andrew argues that fully autonomous inbox-level AI could become very compute-intensive Historical staffing comparison: Maybe 50 people before, now 15–20 could suffice - AI changes execution needs and company org design

Pivotal Quotes: "We don’t think of ourselves that way anymore. We think of ourselves as we are an AI with email features built in." — Andrew Lee: Describing Shortwave’s strategic shift from email client to AI-first communication platform "The thing that matters, the only moat that probably matters is speed." — Andrew Lee: Explaining why Shortwave wants a small, high-velocity team "What if we just run the big LLM a whole bunch of times repeatedly until we get the right answer?" — Andrew Lee: Summarizing the core agent architecture that replaced older brittle heuristics

Implications: For users, AI can now do real inbox work instead of just assisting. For builders, model choice, caching, and iteration matter more than elaborate pipelines. For the industry, software teams may get much smaller, faster, and more agent-managed.