Clay Bavor spent 18 years at Google, including work on Workspace, Google Labs and AR/VR, before co-founding Sierra with Bret Taylor. At Sierra he works on customer-facing AI agents that combine language models with business processes, integrations and guardrails. — Sequoia Training Data Interview.

Visual summary of operating lessons from Clay Bavor.

Part 1: The AI Shift

  1. On the demand for advanced AI: Bavor expects demand for frontier-level AI to remain very large as organizations find more work for increasingly capable systems; this is his forecast, not a measured upper bound. — 20VC Clay Bavor Interview.
  2. On rapid technology adoption: Bavor compares the web’s gradual reach with ChatGPT’s much faster adoption, while he and Taylor discuss how earlier PC, internet and smartphone infrastructure made AI distribution easier. — Acquired Conversation.
  3. On recognizing early capabilities: Bavor recalls an early PaLM test that explained the 2008 financial crisis using an Inception analogy; for him, that ability to connect concepts signaled a new capability. — Sequoia Training Data Interview.
  4. On global model competition: Bavor speculates that distillation of US frontier systems helps explain the strength of some Chinese open-weight models, while US labs may have less incentive to release close substitutes for their paid models. He presents this as an impression, not proven origin for every model. — 20VC Clay Bavor Interview.

Part 2: Engineering Agent Software

  1. On a new category of software: Bavor describes customer-facing agents as a new software building block: probabilistic model outputs must be combined with business rules, tools and safeguards. — Sequoia Training Data Interview.
  2. On fixing AI mistakes: In Sierra’s experience, Bavor says additional AI review can catch some errors that a primary model does not avoid on its first pass. This is an engineering observation, not a guarantee for all models or tasks. — Sequoia Training Data Interview.
  3. On supervisor models: Sierra uses supervisory agents to check primary-agent responses for factual problems, inappropriate advice and prompt-injection attempts before or during customer interactions. — Sequoia Training Data Interview.
  4. On the new development lifecycle: Because model behavior is non-deterministic, Bavor says teams need different methods for programming, testing and deploying agents than for fully rule-based software. — Sequoia Training Data Interview.
  5. On strategic technical depth: Bavor favors building application and agent-system expertise rather than training a general frontier model from scratch; Sierra combines purchased frontier capability with its own architectures and specialized models. — Acquired Conversation.
  6. On internal infrastructure: Sierra’s internal MCP gateway connects approved company systems so agents can work with information a particular employee is permitted to access; Bavor also describes a company-context agent for reasoning across board letters and operating reviews. — 20VC Clay Bavor Interview.

Part 3: Changing the Workforce

  1. On shifting engineering budgets: Bavor expects token allocation to become part of engineering headcount planning, though he says Sierra had not yet reached a formal per-employee budget at the time of the interview. — 20VC Clay Bavor Interview.
  2. On the rising cost of AI tools: Bavor said he had heard of and observed some top engineers using coding agents at more than $100,000 in annualized token spend; he explicitly said Sierra’s own usage was then more modest. — 20VC Clay Bavor Interview.
  3. On the value of young talent: Bavor argues that young people with little formal experience can be unusually valuable when they have strong practical fluency with AI tools; he cites effective Sierra employees in their early twenties. — 20VC Clay Bavor Interview.
  4. On leveling the playing field: At Sierra, Bavor says some recent graduates are more comfortable with AI tools than more experienced colleagues; he does not claim that age alone predicts overall effectiveness. — 20VC Clay Bavor Interview.
  5. On modernizing the technical interview: Sierra asks engineering candidates to build an app using their preferred coding agents and tools, gives them a $150 token allowance, and discusses their architecture and decisions afterward. — 20VC Clay Bavor Interview.

Part 4: Redesigning Customer Interaction

  1. On the primary interface of the future: Bavor predicts that a company’s own AI agent could become a more important customer interface than its website or app; this is a strategic forecast, not a completed shift for all companies. — Acquired Conversation.
  2. On humanizing automated support: Bavor’s customer-experience thesis is that conversational agents can remove some phone-tree friction while still being designed and coached for the brand’s tone and empathy. — Sequoia Training Data Interview.
  3. On extending beyond deflection: Sierra’s Horizon announcement describes agents pursuing multi-step outcomes such as loan origination and sales follow-up, going beyond one-off answers or support deflection. — Sierra: The Next Horizon in Agents.
  4. On capturing brand identity: Bavor says a useful customer agent must reflect the company’s voice, policies and processes; designing that experience requires close work with customer-experience teams, not model engineering alone. — Sequoia Training Data Interview.
  5. On moving beyond cost savings: Bavor frames agents as a potential source of better customer satisfaction, retention and new revenue—not only lower support costs—when they successfully complete customer work. — Sequoia Training Data Interview.

Part 5: Operating at Startup Speed

  1. On embedded customer deployment: Sierra embedded engineers closely with early customers to learn their processes and build the first agent; Bavor treats forward deployment as a useful accelerator for complex enterprise rollouts, not a requirement for every customer. — 20VC Clay Bavor Interview.
  2. On guiding the initial rollout: Bavor says Sierra often guides the first implementation while the customer steers it; he cites Next going live across phone and chat within six weeks from kickoff. — 20VC Clay Bavor Interview.
  3. On aligning governance with technology speed: Sierra holds board meetings roughly every six weeks and sends written memos ahead of them so its directors can challenge current assumptions in a fast-changing AI market. — 20VC Clay Bavor Interview.
  4. On reacting to model upgrades: Bavor says a step change in coding-agent capability after a winter break changed Sierra’s development and product approach; frequent reviews let the company update its assumptions quickly. — 20VC Clay Bavor Interview.
  5. On overcoming early jargon fatigue: Bavor says Sierra tries to explain its products plainly rather than chasing each new AI buzzword as labels shift from machine learning to generative AI to agents. — Acquired Conversation.
  6. On early hardware foundations: In Google’s VR work, Bavor describes improving 360-degree video projection so available pixels were distributed more evenly across the scene instead of wasting resolution near the poles. — Voices of VR Interview.

Part 6: Designing AI Around Outcomes

  1. On matching models to tasks: Sierra composes frontier, open-weight and smaller specialized models, choosing different systems for different steps rather than treating one model as optimal for every task. — Sequoia Training Data Interview.
  2. On fine-tuning where nuance matters: Bavor says Sierra can swap in smaller specialized models for narrow tasks such as triage or planning when they are faster, cheaper or better suited than a broad frontier model. — Sequoia Training Data Interview.
  3. On building above the model layer: Bavor describes the durable application layer as agent architecture, integrations, retrieval, reasoning scaffolds and supervisors built around replaceable foundation models. — Sequoia Training Data Interview.
  4. On pricing delivered value: Bavor says Sierra charges when an agent completes a customer task rather than simply for seats, messages or tokens, aligning its revenue more closely with delivered results. — Acquired Conversation.
  5. On defining success before deployment: Sierra says customers agree on what counts as an outcome before deployment, with transparent criteria for complex cases and escalations rather than an opaque resolution charge. — Sierra: Outcome-Based Pricing.
  6. On choosing a pricing model by job: Sierra says outcome pricing does not fit every interaction; routing or greeting work can instead be billed on a consumption basis within a blended arrangement. — Sierra: Outcome-Based Pricing.

Part 7: Building Systems That Improve

  1. On giving agents proper scaffolding: Bavor and Taylor say reliable agents need a harness of tools, memory, context, a coherent action space, and planning and reasoning capabilities. — Sierra: Agents as a Service.
  2. On designing software for machine users: For software operated by agents rather than people, Bavor and Taylor argue that direct access to data and actions can matter more than menus and forms designed for human clicking. — Sierra: Agents as a Service.
  3. On turning operating knowledge into behavior: Sierra says builders can supply SOPs, support transcripts, whiteboard photos and process notes; Ghostwriter turns those inputs into proposed agent behaviors, integrations and guardrails. — Sierra: Agents as a Service.
  4. On closing the improvement loop: Bavor and Taylor describe a post-launch loop in which agent tools inspect customer interactions, surface weaknesses, validate changes and prepare improvements for review. — Sierra: Agents as a Service.
  5. On pursuing goals across conversations: Bavor and Taylor say long-horizon agents need context and planning across multiple inbound and outbound contacts over days or weeks to complete complex customer outcomes. — Sierra: The Next Horizon in Agents.
  6. On learning from every result: The Horizon announcement says outcomes and rejections feed into an agent’s future decision-making; this is Sierra’s stated product design, not independently measured learning in every deployment. — Sierra: The Next Horizon in Agents.