How I’m Coding in 2026 (The Super-App Strategy)
TL;DR
Every major AI lab is converging on the same app shape — Riley argues Anthropic, OpenAI, Perplexity, Manis, and GenSpark are all building a “super app” with projects and chats on the left, an active agent in the middle, and a preview/browser pane on the right because agentic work rewards fast multitasking across parallel threads.
The real moat is models plus integrations — He says Anthropic and OpenAI have a model advantage, while Google’s Anti-Gravity and Meta-backed Manis may have the strongest integration advantage thanks to Gmail, Docs, Calendar, Facebook, Instagram, WhatsApp, and Ads Manager.
Agentic coding in 2026 looks like managing 5–10 tasks at once — Riley says the best developers he meets in San Francisco are working across multiple parts of a codebase in parallel with agents, using chat switching as the core productivity move while the bots work.
Claude Desktop is really three risk levels in one product — He frames Chat, Co-work, and Code as easy/medium/hard or low/medium/high risk, and personally prefers Claude Code with bypassed permissions because he wants the agent “loose” on his computer instead of stuck in safer sandboxes.
The biggest leverage is not prompting — it’s building your own documentation layer — His practical advice is to spend 3–4 hours writing SOPs, examples, and role docs in Notion or Obsidian so skills/connectors can pull from a single source of truth about your job, goals, and workflows.
Don’t over-optimize tool choice; optimize portability — Riley recommends picking Claude Desktop or Codex and using one daily for a couple months, while storing skills, templates, and context outside the platform so you can move when the leader changes.
The Breakdown
The Big Realization: Everyone Is Building the Same AI App
Riley opens after spending three and a half hours testing Claude Code and comes away with one blunt conclusion: every major AI company is shipping the same product. He points to the mirrored layouts across Claude Desktop, OpenAI’s leaked Codex “super app,” and general agent tools like Perplexity Computer, Manis, and GenSpark — projects and chats on the left, active work in the center, preview on the right.
Why the Interface Keeps Converging
His explanation is simple: agents take time, so people need something to do while they wait. That makes quick chat-switching the killer feature, and Riley says the best developers he knows are juggling five to 10 parallel tasks across a codebase by talking to multiple agents at once.
The Race Comes Down to Models vs. Integrations
Riley breaks the field into two advantages: model quality and integrations. He gives Anthropic and OpenAI a clear model edge, is skeptical Cursor can train a truly top-tier coding model, and notes Claude subsidizes usage heavily — saying $200 in Claude Desktop can translate to roughly $4,000 in token usage, while the same spend disappears fast in Cursor.
The Sleeper Pick: Manis, Plus Google’s Quiet Power
The most interesting dark horse for him is Manis, especially after its Meta acquisition. He says Meta and Google have something the others don’t: giant ecosystems people already live inside, from Gmail, Docs, Sheets, and Calendar to Instagram, WhatsApp, Facebook, and Ads Manager — and he highlights that Manis already lets users do things like control Facebook ads and automate Instagram influencer outreach in ways other agent platforms can’t.
Pick One Tool and Stop Spinning
After surveying the landscape, Riley’s advice is surprisingly non-dramatic: you probably can’t go that wrong. If you’re unsure, he says start with Codex and Claude Desktop because Claude Opus and “Codex 5.4” are, in his view, the best models right now, and then commit to one tool for a few months since they’re all converging anyway.
Inside Claude Desktop: Chat, Co-work, and Code
He then shifts into a hands-on walkthrough and explains Claude Desktop as three modes with different risk levels. Chat is safer, Co-work is sandboxed for document-heavy jobs like law, and Code is the fully unleashed version; Riley admits he runs with bypass permissions on because he’d rather move fast and correct mistakes later than babysit every approval.
Making Agents Autonomous with Scheduling and Remote Control
The practical demo centers on giving agents more independence. He shows Claude Code generating an Anthropic news presentation, then scheduling it to rerun daily at noon, and also demos remote control from his phone by using /remote/control in the terminal so Claude can search his Downloads folder, find a PDF, and summarize it from anywhere as long as his computer stays on.
The Real Work: SOPs, Skills, Memory, and a Portable Knowledge Base
Riley says the highest-leverage thing users should do is stop obsessing over prompts and start documenting their job. He recommends taking three or four hours to map your responsibilities, tools, documents, examples, and goals into Notion or Obsidian, citing his own “master database,” CMO/content creator role docs, and “Paragon templates” based on creators like Callaway.
His Closing Playbook for Building Better Agents
He ends with a six-part checklist: document your role, list your tools, save great examples, write SOPs, identify where you waste time, and define your goals clearly enough that an agent can measure progress. His final point is the most durable one in the video: AI will keep improving, so the safest bet is building clean, portable context outside any one platform and then letting the agents loose.