AI Sprint: Windsurf & Devin — Day 2 (Building Superplexity: Plan with Opus, Execute with Cascade, Review with Devin)

The taxonomy: an IDE (Windsurf, Cursor, VS Code descendants) is where YOU set up the codebase and collaborate with an agent locally; an AUTONOMOUS CODING AGENT (Devin) takes a task — 'fix the bug, implement the feature' — into its own CLOUD SANDBOX and works it end to end. Sandbox gets the restaurant treatment: 'instead of testing the wild dish in the main kitchen, the chef says: here is a testing kitchen. Make it there, taste it, then bring it in' — a protected environment where experiments can't hurt production. The pairing insight: Cursor+Claude Code is 'exactly like' Windsurf+Devin — IDE plus agent — and because every IDE descends from VS Code, the official Claude Code extension runs inside Windsurf too: 'any IDE you think of, there's 70-80% feature overlap. I can combine anything in any way I want.'

Devin's differentiators beyond the sandbox: Slack integration, automatic PR answering, and persistence — plus enterprise-shaped pricing ('a little off the curve') and a focus 'purely on software... not optimized for non-coding tasks,' where Claude Code and Codex stay the all-rounders.

The ritual, demonstrated end to end: open Claude Code INSIDE Windsurf (VS Code extension), invoke the Obra SUPERPOWERS brainstorming skill, and let it interview you into a spec — 'I'm not jumping in with my features; I'm setting up an interview between me and Claude Code.' Opus plans because 'Claude Opus is one of the best planning things available.' Then the artifact chain: the BRAINSTORM produces a SPEC ('proper English — for YOU to verify'), the writing-plans skill converts it to a PLAN ('very technical — for the AGENT... user stories, sprints, all written in'), and Cascade implements it by file reference — '@ the plan file, say implement this' — kicking off 'sub-agent-driven development' on Windsurf's cheap adaptive models. No copy-paste between tools: Claude wrote docs/plans/ into the same folder Cascade reads.

The non-technical escape hatches are taught as first-class moves: 'I am a completely non-tech person — simplify this and tell me where you need my direction' resets a spec gone jargon-heavy; recommendations can be overridden ('you can push back — I don't want this, do that'); and skills install by pasting the GitHub link ('drop it into Claude Code and say install it — don't have to worry').

The interview surfaces every decision Day 1 said a spec needs. SOURCES: Substack/Beehiiv newsletters and curated Twitter accounts as the high-signal pool, hybrid with wider search; the source LISTS are agent-researched ('get me the top AI substacks, beehiivs, accounts — search yourself'). SCORING: weighted blend — all sources run in parallel, rescored by source weight (newsletters 1.0, tweets 0.7, Google fallback 0.4) — chosen over pure tiering or classifier routing. COST CONTROL: Twitter via Apify actors ('the X API is $200/month; scraping is fragile — hybrid recommended'), and the hard rule PAID SCRAPING ON DEMAND ONLY — 'only when I ask, with that scope. Don't keep scraping; my money will get drained' — with the clarifying nuance the agent itself surfaces: free RSS polling can stay background; paid calls cannot. EDGE CASES: six defaults accepted explicitly — thin results ('synthesize and SAY thin coverage'), paywalls (cite what's accessible, flag it), conflicting sources (surface the disagreement), hallucination post-check, graceful API-failure degradation, stale-list passive feedback. SEARCH: Sonar (Perplexity's model, via OpenRouter) replaces Google entirely — 'I'm building a better Perplexity, so I'm using Sonar.' STACK (recommended, accepted): thin orchestrator, SQLite FTS5, no vector store ('top 20 articles back to Sonnet — we don't need one'), Next.js front end, keys in .env.

The demo: point Devin at the model-council repo (built in an earlier sprint) — 'go through the repository, tell me if there are bugs, discuss with me before next steps.' Devin clones it into its cloud sandbox and returns a structured REVIEW REPORT: 4 bugs (duplicate prompts wasting tokens, an .env sentinel mismatch), 7 code-quality issues, missing features, improvement ideas. On 'go ahead and fix,' it creates 15 tasks, works its OWN BRANCH, raises a pull request, runs the CI checks, ANALYZES ITS OWN PR for merge-worthiness ('if everything is fine, it merges; if not, it won't') — and finds additional bugs mid-fix, pushing those too. The GitHub view closes the loop: two branches, the PR trail, automated checks — Day 1's pizza mechanics, operated by an agent.

The context that frames trust: Cognition's founder Scott Wu — three IOI golds, ICPC podium, 'legendary grandmaster... one of the best coders in the world is building the software-engineering agent.' And the lane discipline: 'more for software engineers, but non-engineers can use it — it's just more technical'; enterprise pricing; for greenfield, Windsurf; for legacy, Devin.

The failure chain, kept in: the app runs, the first research query returns the thin-coverage default (working as specced!) because the DATABASE WAS EMPTY — no newsletters indexed yet. The fix is the standing pattern: copy the error, paste it to the executor ('this is the error I got — can you check?'), let it index the sources and retry. The working answer that follows has its own honest wrinkle: the sources are 'a little old — it says 4.6 while we are at 4.7' — retrieval freshness as a visible product property. His verdict calibrates expectations: 'still some improvements, but a decent start... a minimum viable product. It's in markdown — you'll have to say render it properly. Keep giving it feedback and improve it further.' And the edge-case humility: 'ideally the edge case should have caught that, but some slip through the cracks. You go back and fix.'