How I AIHow Devin replaces your junior engineers with infinite AI interns that never sleep | Scott Wu (CEO)
CHAPTERS
- 0:00 – 3:07
Devin as an async “junior engineer”: task-focused automation
Scott frames Devin as a junior engineer that shines on well-scoped tasks rather than open-ended problems. Claire reinforces the “infinite interns” mental model and sets up the episode’s focus on practical workflows and demos.
- •Devin is best at tasks, not broad architectural problems
- •Intern metaphor: delegate, then check in later
- •Success depends on clear scoping and handoff
- •Sets expectation for async collaboration and multiple parallel tasks
- 3:07 – 4:04
The Devin stack: Slack, Linear, DeepWiki, Ask Devin, and PR-based delivery
Scott outlines the different surfaces where Devin lives and why that matters for real teams. The conversation emphasizes that Devin behaves less like an IDE feature and more like a teammate embedded in existing workflows.
- •Key components: Slack/Linear integrations, DeepWiki, Ask Devin, Devin sessions
- •Devin is positioned as an “employee” rather than a tool pane
- •Choosing the right surface depends on role (IC vs lead/PM)
- •PR output is the natural artifact for review and merge
- 4:04 – 7:01
Where Devin excels: backlog, engineering toil, and repeatable execution
Scott lists the recurring categories where Devin reliably adds value: triaging backlog items, handling toil, and responding to operational issues. The focus is on high-leverage, repeatable work that benefits from automation.
- •Issue backlog triage and first-pass fixes
- •Engineering toil: version upgrades, docs, unit tests
- •Diagnosing crash reports and operational debugging
- •Tag Devin early to accelerate throughput
- 7:01 – 9:10
DeepWiki walkthrough: understanding the codebase with English + code context
Scott demonstrates DeepWiki as an AI-generated documentation layer over a repo, showing how it helps locate the right files and patterns. Claire notes the value of pairing natural-language explanations with exact code references for trustworthy understanding.
- •DeepWiki generates structured, searchable repo documentation
- •Combines narrative explanation with code snippets and file pointers
- •Used to locate MCP marketplace/server list implementation
- •Helps build accurate mental models before delegating work
- 9:10 – 11:24
From research to execution: turning context into a Devin-ready prompt
They show how DeepWiki can transform a vague request into an implementation-ready prompt that references the right files, types, and patterns. Claire calls out the “prompt improvement loop” as a key habit that reduces rework.
- •Start with an intent (“add ChatPRD MCP server”), then expand with repo context
- •Prompt includes: existing examples (Atlassian/HubSpot), types, functions, checks
- •Better prompts reduce async failure and back-and-forth
- •DeepWiki-to-Devin handoff becomes a repeatable workflow
- 11:24 – 15:42
Async collaboration model: sync setup, async work, and multithreading tasks
Scott explains why async is central: once a session starts, Devin researches, edits code, and prepares a PR without needing constant supervision. Claire highlights how async design reduces latency frustration and enables running many tasks in parallel.
- •Two phases: quick synchronous alignment, then asynchronous execution
- •Devin researches externally and in-repo while you do other work/meetings
- •Parallelize: run multiple Devin sessions like multiple interns
- •Outputs include PR diffs and (for UI work) before/after screenshots
- 15:42 – 19:19
Live demo friction: when the agent needs a pointer (and how to unblock it)
During the MCP integration demo, Devin struggles to find the right documentation, and Claire supplies the exact URL. They treat this as a realistic “intern asks for help” moment and reinforce that occasional guidance is expected.
- •Agents can miss key resources during web research
- •Humans can unblock quickly by providing the right link/artifact
- •Reinforces intern metaphor: async work + periodic check-ins
- •Goal remains a reviewable PR for the integration
- 19:19 – 21:28
Institutionalizing Devin in Slack: first-line response and PR-centric iteration
Scott shows how Cognition operationalizes Devin by tagging it as the default first responder in issue channels. Examples illustrate a loop of request → Devin session → PR → feedback → additional commits → merge.
- •Devin as “first-line of response” across Slack issue channels
- •Examples: UI typography standardization; cmd-click notification open in new tab
- •Devin posts progress updates, files inspected, and confidence level
- •Human review and small touch-ups remain part of the workflow
- 21:28 – 23:00
Multiplayer agent threads: humans collaborate around Devin’s work
They emphasize that Devin sessions are naturally collaborative—multiple engineers can add guidance in the same thread. This turns agent work into a shared team activity and accelerates convergence on the correct implementation pattern.
- •Multiple humans can coach Devin in one Slack thread
- •Experts provide precise implementation guidance (e.g., TanStack Router link element)
- •Devin incorporates feedback via incremental commits
- •Agent work becomes visible and auditable through the thread + PR history
- 23:00 – 25:49
Driving AI adoption: do it in public channels to upskill both agent and team
Claire argues that public usage increases adoption and teaches people how to prompt effectively by observation. Scott adds that public work benefits both sides: Devin’s institutional knowledge improves, and humans learn patterns and best practices from each other.
- •Public agent threads improve team learning and normalize usage
- •Shared visibility improves prompting skills and expectation-setting
- •Knowledge compounds: Devin reuses context from prior work
- •Trend toward multiplayer AI productivity experiences
- 25:49 – 32:27
Top engineering tasks for Devin: the “top five” playbook
Scott lists his top five high-ROI task categories, and Claire adds her own overlapping set plus an operational workflow for auto-updating docs. The segment becomes a pragmatic checklist for teams adopting coding agents.
- •Scott’s five: frontend fixes, version upgrades/migrations, documentation, incident response, adding tests
- •Claire’s additions: UI polish/magical moments; Devin-generated PR descriptions; auto-shipping internal docs
- •Sentry/PagerDuty-style integrations for first-line incident handling
- •“24/7 rubber duck” support as a personal productivity booster
- 32:27 – 36:07
ChatGPT Voice in meetings: adding an AI participant for real-time research
Scott describes using ChatGPT voice to answer meeting questions without disengaging or silently Googling. Claire reframes voice as socially inclusive when treated as a shared participant whose answers everyone can hear in real time.
- •Voice lowers friction vs typing; faster than “traditional Google” lookup
- •Useful for quick, off-the-cuff factual research during meetings
- •Avoids the social cost of phone/laptop distraction
- •Shared audio response keeps the whole room aligned
- 36:07 – 38:51
Lightning round: the future interface—agents, IDEs, Slack, and beyond
Scott predicts the longer-term future is an agentic human-computer interface where you manipulate the product directly while an agent implements changes. Near term, he expects different entry points depending on role: Slack/Linear for leads and the IDE for ICs.
- •Long-term: agent as next-gen interface ("Jarvis" model), product-first manipulation
- •Near-term: Slack/Linear workflows fit meeting-heavy leads; IDE fits coding-focused ICs
- •Form factors will be flexible and adapt to user context
- •Agents run background tasks while humans stay in their preferred surface
- 38:51 – 41:11
Prompting when stuck: inspect the agent’s history and correct the missing step
Scott’s core troubleshooting advice is to treat agent work like pair debugging: review logs to see what the agent did, identify the missing file/link/assumption, and then re-instruct with that specific information. The episode closes with where to find Devin/Cognition and the wrap-up.
- •Agents differ from chatbots because you can audit their action history
- •Use logs to find the failure point (missing link, missed file, wrong assumption)
- •Re-instruct with targeted context to unblock and continue
- •Wrap-up: Cognition/Devin on X; devin.ai for the product
