Up your game with async agents w/ Jessie Frazelle
In this episode, we talk with Jessie Frazelle (Zoo, Oxide, ACM Queue) about what it takes to get real value out of async AI agents in your dev workflow. We covered: - Why most AI code review agents add noise—or worse, false confidence - The distinction between agents that summarize vs. agents that spot real bugs - How async agents shift from "helpful suggestions" to "actual second opinions" - Lessons from shipping code across Rust, TypeScript, and C++ with async agents watching - What makes a high-signal bot: sparse comments, zero hype, and earned trust - The line between tooling you use vs. tooling you start depending on Jessie walks through experiments across dozens of tools: where they failed, what stuck, and why one async agent quietly earned a spot in her team’s stack.
Video Chapters
- 0:00 Begin your journey: Synchronous vs. Asynchronous AI agents unpacked
- 2:36 The early struggle: When AI code summaries fell short
- 4:35 To comment or not to comment: Defining AI's true value
- 9:40 Beyond replacement: AI's true role in augmenting human review
- 13:41 A game-changer appears: Discovering the power of Diamond-Graphite
- 15:06 Catching the uncatchable: Diamond identifies a subtle math error
- 18:12 From math to logic: Diamond's impact on critical code flaws
- 20:47 The true strength: AI as your ultimate second pair of eyes
- 23:40 The core promise: Finding the bugs humans just can't see
- 25:56 The golden rule: Why low-touch, high-value AI wins
Original Output
0:00 Begin your journey: Synchronous vs. Asynchronous AI agents unpacked 2:36 The early struggle: When AI code summaries fell short 4:35 To comment or not to comment: Defining AI's true value 9:40 Beyond replacement: AI's true role in augmenting human review 13:41 A game-changer appears: Discovering the power of Diamond-Graphite 15:06 Catching the uncatchable: Diamond identifies a subtle math error 18:12 From math to logic: Diamond's impact on critical code flaws 20:47 The true strength: AI as your ultimate second pair of eyes 23:40 The core promise: Finding the bugs humans just can't see 25:56 The golden rule: Why low-touch, high-value AI wins Timestamps by StampBot 🤖
Unprocessed Timestamp Content
0:00 Starting off with synchronous versus asynchronous AI agents 0:59 Initial experiences with PR review bots, mostly low-value 1:51 Overly verbose AI reviews and the questionable value of summaries 2:36 The unreliability of AI agents generating accurate code summaries 3:26 Unnecessary bot comments cluttering pull requests: just don't! 4:05 Coderabbit-ai's confusing comments on a simple readme change 4:35 The debate: should bots always comment or only add value? 5:04 Jessie explains the annoyance of constant low-value bot chatter 6:19 Graphite's CI integration and helpful, non-intrusive feedback 7:10 Greptile-apps bot provides a summary and simple "LGTM" 8:00 More Copilot examples: often unhelpful and easily ignored 9:00 CodiumAI-Agent's review guide: vague, verbose, low-value text 9:40 AI agents aiming to replace, not augment, human code review 11:05 Early attempts to increase test coverage using Codex bot failed 12:15 Human developers find bot-generated tests often lacking quality 13:06 Codex-generated pull requests for reducing API data size 13:41 Diamond-Graphite emerges as a surprisingly useful AI tool 14:24 Diamond bot finding subtle mathematical inconsistencies in code 15:06 Specific example of Diamond bot catching a real math bug 16:29 Diamond's ability to learn custom languages and find errors 17:24 The positive impact of Diamond: focused, clear, and valuable suggestions 18:12 Real-world examples of Diamond catching critical logic bugs 18:56 Jessie and her colleague's humorous attempts to troll the bot 19:27 The bot flags "non-serious feedback," proving its sensitivity 20:47 The actual value of AI: a second pair of eyes, not a replacement 21:55 The dilemma of blind acceptance versus verifying bot suggestions 22:50 Diamond bot excels at finding deeply hidden mathematical errors 23:40 The core value: a bot finding bugs humans would otherwise miss 24:20 Overview of the diverse tech stack and AI tools utilized 24:50 Gemini, Claude, and general AI capabilities for coding 25:56 Final takeaway: low-touch, high-value AI is the sweet spot Timestamps by StampBot 🤖