CHAPTERS
- 0:00 – 4:53
Why the future is cloud agents—and why “agent management” becomes the real job
Ryan argues that local-only engineering is becoming obsolete as cloud agents take over continuous execution. He frames the core skill as managing many agents like an organization manages many employees, including the overhead of scale and delegation.
- •Cloud agents enable 24/7 work without being tied to a laptop
- •Ryan’s spend on Devin illustrates how quickly usage scales with reliance
- •The bottleneck shifts from coding to coordinating and supervising agents
- •Agent management mirrors classic org design (limits on direct reports)
- 4:53 – 8:17
Ryan’s Untangle journey: divorce assistant → law-firm software (PMF via B2B pivot)
Ryan shares how Untangle began as a consumer AI divorce product but failed to find adoption. A meeting with a family lawyer revealed a clear B2B use case (paralegal shortage, discovery pain), unlocking willingness to pay and rapid growth.
- •Initial consumer positioning didn’t work: users still want humans/lawyers
- •A single lawyer conversation reframed the product as a firm tool
- •B2B value props: discovery workflow, staffing shortages, operational relief
- •Success created “max Q” pressure: solo founder overwhelmed, hiring imminent
- 8:17 – 16:28
Ryan’s current Devin stack: folders, priority triage, and the weekly paper list
Ryan walks through how he structures Devin work using folders (Bugs, P0/P1/P2, investors) and limits active concurrency. He pairs this with a paper-based weekly priority list to stay anchored amid many threads and screens.
- •Devin becomes the primary workspace: many concurrent threads
- •Foldering by priority (P0/P1/P2) and category reduces cognitive load
- •Physical paper weekly priorities help prevent distraction and drift
- •Operational tip: build repeatable playbooks/skills like investor updates
- 16:28 – 16:29
Watchdog playbook: monitoring every customer account like an always-on ops console
To manage customer success and engineering simultaneously, Ryan built a Watchdog playbook that inspects each firm’s account activity and surfaces the most important issues. It summarizes errors and status (fixed, in progress, or stuck) to guide next actions.
- •Watchdog scans each customer/firm: activity, errors, and UX friction
- •Transforms raw telemetry into prioritized “top three” problems
- •Checks whether issues are already fixed or blocked by unmerged PRs
- •Used as a stress-release tool when Ryan feels blind to business status
- 16:29 – 19:10
Decision fatigue and focus: task buckets vs. time buckets (Ryan vs. Claire)
Ryan and Claire compare how they stay effective amid constant agent output. Claire buckets her time rather than tasks and uses “what dies on the vine” as a prioritization signal, while Ryan relies on structured folders and playbooks.
- •Claire: time-blocking for co-piloting vs. spinning off/checking in
- •Claire: decision fatigue becomes a heuristic for what not to pursue
- •Ryan: structured task triage with folders + repeatable playbooks
- •Both acknowledge agents enable longer workdays and more decisions
- 19:10 – 22:51
More code ≠ better product: the anti-auto-improvement reality and the need for customer conversations
They caution that scaling output doesn’t automatically create a better product because models can’t reliably decide what to ship. The way Ryan found PMF was classic: outbound, calls, and sitting with customers—work that AI doesn’t replace.
- •Frontier models still can’t choose product direction reliably
- •Automatic “improvement loops” fail in product because markets don’t appear magically
- •Mismatch: massive code output vs. limited commercial demand
- •Ryan’s PMF method: outbound email → customer calls → on-site observation
- 22:51 – 24:26
Cloud agents beyond code: ops, deal desk, documentation, and customer triage
Claire highlights that cloud coding agents are underused when treated as code-only tools. Because they can both understand the codebase and execute changes, they can run operational workflows like deal desk, documentation, and support triage.
- •Background agents can operate across the business, not just engineering
- •Examples: deal desk, ops automation, docs generation, customer triage
- •“What would you do with someone who knows your codebase and can fix anything?”
- •Trend: coding platforms expanding toward general business agents
- 24:26 – 26:54
When to use Codex vs. Devin: low-latency pair programming vs. autonomous background work
Ryan and Claire compare tool choice by workload type. Codex excels for interactive, UI-heavy grinding and verification due to low latency and local browser access, while Devin shines for cloud-based async execution and playbooks.
- •Codex: best for pair programming, front-end/UI iteration, long refactors
- •Codex: strong for verification loops (CLI checks, browser-based user-story testing)
- •Devin: strong for background execution, quick PRs, and cloud autonomy
- •Choosing tools is about latency, context, and supervision needs
- 26:54 – 31:17
PR throughput systems: Merge Mommy (risk-based auto-merge) vs. Ryan’s LAN PR playbook
Claire explains Merge Mommy, an automated PR risk assessor that auto-approves low-risk changes and escalates medium/high risk for human review. Ryan describes LAN PR: Devin review loops, recorded video walkthroughs, and merge-queue discipline to land PRs safely.
- •Merge Mommy: triggered after CI; scores PR risk; auto-approves low-risk
- •Compliance-driven workflow: higher risk PRs ping humans for review
- •LAN PR: runs Devin Review, iterates up to two review/fix loops
- •LAN PR: records narrated video walkthrough; merge only after “video approved”
- 31:17 – 35:18
Slack vs. Devin threads: where teamwork and context should live in the AI age
Ryan argues for moving collaboration into Devin threads to keep context unified and reduce Slack-driven noise, even suggesting banning DMs. Claire defends Slack as a change-management layer for making agent work visible, while agreeing Slack culture often dilutes execution.
- •Ryan: Slack fragments context; prefers discussion inside agent threads
- •Ryan: anti-DM stance; wants conversations tied directly to work artifacts
- •Claire: Slack made early agent adoption effective via public, real-time visibility
- •Shared view: reduce “talking that doesn’t result in work” and tighten loops
- 35:18 – 39:04
Design systems workflow: Claude Design for tokens, then Codex for implementation
Ryan says Claude Design is uniquely good at carrying forward a design system, while Devin/Codex lag on design consistency. Claire shares a pipeline: generate a design.md and token set in Claude, then hand it to Codex to build a technical component system and connect it to the repo and Figma.
- •Claude Design excels at extracting/decomposing design systems into reusable tokens
- •Codex performs better at implementing the technical design system in a codebase
- •Practical handoff: Claude-produced design.md → Codex builds components/monorepo links
- •They note frustration with Claude’s “output style” despite its strengths
- 39:04 – 40:05
EA tooling and inbox automation: OpenClaw vs. Claude Code vs. Codex on local hardware
They compare executive-assistant setups for running agent work beyond coding. Claire still uses Polly the OpenClaw, while Ryan moved to Claude Code on a Mac Mini and is considering switching to Codex; both highlight email triage as a major unlock.
- •Claire: OpenClaw remains her EA backbone; also uses Codex for email workflows
- •Ryan: shifted off OpenClaw to Claude Code on Mac Mini; prefers Codex UI
- •Codex as an inbox operator: processing thousands of unread emails quickly
- •EA role becomes orchestrating and delegating across multiple agent surfaces
- 40:05 – 42:09
Hiring in the agent era: evaluate candidates by how they manage agents (with recordings)
Ryan describes a hiring process optimized for AI-native execution: candidates submit full-screen recordings of building features, prioritizing agent-management skill over interview charisma. Next phases include giving Devin access to build real work and using output visibility (including Devin replays) for evaluation.
- •Screen-recorded builds reveal real working style and agent-management capability
- •Phase-based funnel: demo feature build → real Devin task → short human chat
- •Agent-era hiring emphasizes verification loops over “do I like you?” interviews
- •Devin’s built-in replay reduces overhead and increases transparency
- 42:09 – 44:13
Final recap: ship responsibly, talk to customers, and treat agent management as a core competency
Claire summarizes the episode’s paradox: the most AI-forward founders still stress that product quality and PMF require human understanding, not just code volume. They close with practical takeaways on prioritization systems, expanding agent use beyond code, and the reality that both are now hiring humans.
- •Too much AI output doesn’t guarantee product-market fit
- •Customer conversations remain the core differentiator
- •Use simple prioritization anchors (folders, time buckets, even paper)
- •Everyone becomes an agent manager; tool choice and workflows matter
