Aakash GuptaComplete Course: AI Agent Products (with Warp.dev CEO Zach Lloyd)
CHAPTERS
- 0:00 – 2:51
Warp’s breakout growth: $1M ARR every 10 days and 700k active developers
Aakash and Zach open with the headline metrics behind Warp’s surge: rapid ARR expansion and a large, fast-growing active developer base. They frame AI agents—especially coding agents—as the defining product trend and set up the episode as a tactical playbook.
- •Warp’s monetization acceleration: adding $1M+ ARR every 10 days
- •~700,000 active developers using Warp
- •19X+ revenue growth in a year
- •AI coding agents positioned as the #1 product trend
- 2:51 – 4:19
The inflection point: repositioning from “terminal” to “agentic development environment”
Zach explains that Warp’s turning point wasn’t just adding AI—it was finding the right interface for AI inside the product. Going all-in on agentic features and making them the front door (including a June launch) drove the step-change in growth.
- •Early Warp: reimagining the terminal UX for developer workflows
- •CLI-based agents (e.g., Claude Code, Gemini CLI) validated the interface direction
- •Key decision: make agents front-and-center, not a side feature
- •Repositioning and launch timing correlated with growth acceleration
- 4:19 – 6:00
Why “agentic AI” is now a must-have product primitive
Zach argues product builders are at a “learn it or get left behind” moment because intelligence has become a new foundational building block—like databases or APIs. He emphasizes that most software problems can benefit from embedded intelligence, regardless of category.
- •Intelligence as a new primitive in the product stack
- •Applies across productivity, consumer, and go-to-market tools
- •Agentic features are becoming table stakes for modern apps
- •Builders must proactively explore where intelligence fits
- 6:00 – 8:29
Framework: avoid gimmicks by starting with the user problem (not the model)
Zach lays out a classic product-first approach: define the real user problem, then hypothesize where intelligence can compress time or complexity in a workflow. He illustrates how LLMs replace brittle rule-based algorithms by using context to produce useful outputs.
- •Start with a meaningful customer problem and a clear workflow
- •Form a hypothesis: where does intelligence make the workflow faster/better?
- •LLMs can replace complex rule systems with contextual reasoning
- •Look for tasks where users must express complexity (rules, formulas, code)
- 8:29 – 12:39
Warp’s AI evolution: from English-to-command to chat panel to native “agent mode”
Warp iterated through multiple AI integrations that mirrored the broader market: translation features, then a ChatGPT-like panel, then a more native agent experience. The key UX unlock was using the command line’s natural “tell the computer what to do” pattern, but in English, enabling a tool-using agent.
- •Early feature: translate English into terminal commands
- •Chat panel experiment felt bolt-on and non-native
- •Breakthrough: type English in the same input used for commands
- •Agent mode became a foundation for adding tools (edit files, read web, etc.)
- 12:39 – 19:38
Agent UX principles: don’t bolt on chat—embed intent into the native UI
Zach generalizes the Warp lesson: chat overlays are a thin moat and often the wrong UX. The best agent experiences let users express intent in the app’s native objects (cells, commands, etc.) and leverage agents as “virtual workers” using the product’s tools.
- •Bad pattern: “put chat in my app” as the primary integration
- •Good pattern: agents operate through the app’s native UX primitives
- •Spreadsheet example: AI populating/transforming cells vs. corner chat button
- •Think: what would a human assistant do, then enable the agent to do that work
- 19:38 – 25:18
Live demo: exploring a repo and shipping a real UI change with Warp’s agent
Zach demonstrates Warp’s agent in practice—first to understand a repository, then to implement a tooltip feature in Warp itself. He shows how adding context (screenshots, file references) and iterating with the agent can produce a working code change in a large Rust codebase.
- •Ask natural-language questions to explore an unfamiliar repo
- •Build a feature via a single prompt: tooltip for directory picker chip
- •Provide context via screenshot + file references for better results
- •Agent edits code and builds the project; user reviews and iterates
- 25:18 – 31:05
Making agents feel personalized: rules, memory, and reducing repetition
The conversation shifts to how agent products should remember user preferences and avoid forcing repeated instructions. Zach explains Warp’s approach using persistent “Rules,” reusable prompts, and product design that anticipates the next likely step (like building after a code change).
- •Key UX goal: minimize repetitive instruction from users
- •Rules as persistent context (e.g., ask before committing, always format)
- •Reusable prompts and learned workflows improve speed and stickiness
- •Balancing autonomy with safety/explicit confirmation
- 31:05 – 35:18
How Warp dogfoods: a “start with a prompt” engineering mandate + feedback loops
Zach describes internal adoption systems: engineers begin tasks with an agent prompt and share feedback when it fails or succeeds. Warp collects learnings via dedicated Slack channels and highlights wins (“Warped It”) to educate the team and create external content.
- •Dogfooding as a core mechanism for product improvement
- •Mandate: start every task with a prompt; share failure feedback
- •Dedicated channels to categorize friction and patterns
- •Success artifacts (Looms) become training + marketing content
- 35:18 – 39:45
Competitive landscape: IDE forks vs pure CLI tools—and Warp’s “third category” bet
Zach compares Warp to IDE-based competitors (Cursor/Windsurf-like) and CLI-only agents (Claude Code-style). Warp aims to be an agentic development environment designed around the new workflow: prompt-first, diff visibility, and fast human-agent iteration, rather than a bolt-on chat inside legacy UI.
- •IDE tools: code-first UI with chat as a sidecar experience
- •CLI agents: similar interaction model but less rich UI/controls
- •Warp’s differentiation: purpose-built UX for agentic workflows
- •Trade-offs acknowledged, but category creation is the strategy
- 39:45 – 46:15
Activation and onboarding: trigger agent help at the moment of need (agent “autocomplete”)
Zach explains what didn’t work (guided tours, welcome screens, copy tweaks) and what did: contextual suggestions when users hit errors or friction. Warp predicts the next best action, drops users directly into the agent flow, and delivers an immediate “it solved my problem” aha moment.
- •Ineffective: traditional tours, onboarding modals, button placement tweaks
- •Effective: just-in-time help when users encounter real friction
- •Predict next steps and offer a one-click path into agent mode
- •Model this like “autocomplete for agents”: low-friction, high-upside assistance
- 46:15 – 53:00
Measuring success: depth of engagement, retention smiles, and agent-specific eval rigor
Zach outlines the metrics Warp tracks—especially deep, frequent agent sessions as a leading indicator of conversion. He then details why evals matter for nondeterministic systems, covering public benchmarks, internal task harnesses, and production feedback loops from real user interactions.
- •Primary product signal: depth of agent engagement (long, frequent tasks)
- •Cohort retention and reactivation (“smile” curves) show increasing value
- •Evals are mandatory due to nondeterminism; early versions shipped before evals
- •Use public benchmarks + internal harnesses + anonymized real-world failure analysis
- 53:00 – 1:02:09
Monetization realities: why seat-based SaaS pricing breaks for agents
The conversation turns to business models: agents are costly to run, usage is highly variable, and fixed per-seat pricing misaligns costs and revenue. Zach discusses hybrid pricing (subscription plus overages) and argues outcome-based pricing works best when value is directly measurable (e.g., ticket resolution).
- •Preconditions for success: willingness to pay, retention, and margins
- •Seat-based pricing misaligns incentives when usage drives costs
- •Hybrid approach: fixed subscription + usage-based overages
- •Outcome-based pricing is compelling where value is measurable (hard in coding)
- 1:02:09 – 1:11:31
Where agents are headed: from autocomplete to interactive agents to partial automation + a 90-day PM plan
Zach proposes a three-phase evolution: autocomplete, interactive human-orchestrated agents (today), and eventual automation of simpler tasks. He closes with practical advice for PMs: get hands-on with an agent tool, prototype ideas directly, and build intuition for what models can and can’t do.
- •Three phases: autocomplete → interactive agents → automated task execution
- •Near-term prediction: most dev tasks will start with a prompt
- •Adoption is still early outside tech bubbles despite rapid capability gains
- •PM roadmap: use agent tools daily, prototype ideas, calibrate feasibility through practice
