Silicon Valley GirlHead of ChatGPT & Codex: agents for normal people are HERE
CHAPTERS
- 0:00 – 0:51
Personal AI assistants become default (even for non-prompt power users)
Thibault predicts a near-term shift where everyone effectively gets a “personal assistant” on their computer, regardless of how much they’ve learned prompting. The value won’t come from clever prompts but from agents that proactively help with everyday work.
- •AI benefits will reach non-technical users without active prompting
- •Agents will feel like personal assistants embedded into daily computing
- •The biggest change is usability and reliability, not people’s behavior
- •Creatives and idea-driven builders are positioned to experiment quickly
- 0:51 – 2:05
How knowledge work changes as agents become reliable across tools
Marina frames the transformation as moving beyond software engineering into all knowledge work in the next months. Thibault explains this is driven by agents that can operate over longer horizons and use many tools (browser/computer + integrations).
- •Knowledge work impact will accelerate as agent reliability improves
- •Agents can use browser/computer and many plugins/integrations
- •Past friction required technical users to intervene; that’s fading
- •Mainstream adoption comes from “it just works” execution
- 2:05 – 3:02
Agentic workflow breakthrough: scheduled automation + daily briefs
Thibault describes common automations like recurring market research, summarizing inbound messages, and even printing daily digests. Marina highlights the emerging approval-based “dashboard” where you review what agents did overnight and greenlight next steps.
- •Recurring tasks can run on schedules (e.g., every 12 hours)
- •Agents can summarize Slack/news/email into consumable digests
- •Approval-based workflows: review/approve actions taken overnight
- •Packaging and UX are the remaining barriers more than capability
- 3:02 – 4:18
Safety & trust layer: Auto Review and longer autonomous runs
They discuss how autonomy depends on safety systems. Thibault explains Auto Review: a second agent verifies the first agent’s actions to reduce harmful or risky outcomes, enabling longer, more independent operation on sensitive data.
- •Auto Review: a second agent validates actions from the main agent
- •Designed to prevent harmful mistakes (e.g., leaking personal info)
- •Alignment/safety research enables increased autonomy
- •Trust increases willingness to grant agents broader access
- 4:18 – 4:58
Must-have agent files: examples beat explanations (tone, projects, contacts)
Marina asks how to organize data so an agent can work well. Thibault shares his preference for tidy folders/notes (increasingly moving to cloud), and recommends providing examples of writing rather than trying to “describe” a tone.
- •Maintain structured project folders and notes (agent can help organize)
- •Tone of voice: include real examples (newsletters, messages), not prose rules
- •Contacts and project context are high-leverage inputs
- •Agents can also pull context from existing productivity apps
- 4:58 – 7:07
Cloud memory is coming: one agent across devices (laptop, phone, travel)
Marina worries about working across multiple Macs and travel devices. Thibault predicts a near-term shift to cloud-hosted files and agent memory so users aren’t stuck managing fragmented, device-bound assistants.
- •Local-only workflows create fragmented ‘multiple agents’ across devices
- •Next phase: cloud-hosted memory and files managed by the agent
- •Current solutions are improvised (e.g., Google Drive folders)
- •Goal: consistent identity/context across phone and computers
- 7:07 – 9:28
Use agents vs. don’t: the widening productivity gap—and human responsibility
They explore how people who adapt will outperform those who don’t, but also the burden of oversight. Thibault emphasizes humans remain accountable (as with AI-written code): you can’t outsource understanding or responsibility.
- •Adopters gain leverage on deferred tasks and personal admin
- •Agents may help with taxes, inbox rules, staying in touch with loved ones
- •Humans remain responsible for outcomes (code and non-code alike)
- •Oversight/review is part of effective agent use
- 9:28 – 11:49
The optimization trap: doing everything too early on the capability curve
Marina describes feeling overwhelmed trying to optimize everything with agents. Thibault warns that pushing too far can backfire when capabilities aren’t yet reliable—though experimentation helps map what will become feasible soon.
- •Over-automation can create cognitive overload
- •Some workflows are possible soon, but not reliably today
- •Experimentation reveals boundaries of current reliability
- •Revisit ambitious automations as models improve in 3–6 months
- 11:49 – 13:38
Vibe coding reality check: prototypes vs. scalable products
Marina shares her team’s experience vibe coding small apps that work but don’t scale well architecturally. Thibault advises solo vibe coding is great for experiments, but scaling still benefits from technical expertise—at least for now.
- •Vibe coding works for quick tests and personal/limited sharing
- •Scaling to large user bases still needs engineering rigor
- •Agents will improve at long-term maintainability, but not fully there yet
- •Near-term gains expected in maintainability over 6–9 months
- 13:38 – 15:07
Future of software engineering: more software, more demand, more experimentation
They discuss whether AI reduces the need for engineers. Thibault predicts an explosion of apps and infrastructure because turning ideas into prototypes becomes trivial, expanding the universe of solvable problems and sustaining demand for technical talent.
- •Lower barrier to building increases total software created
- •Creatives can prototype quickly and iterate toward real products
- •Hard to cap demand—new problems and solutions keep emerging
- •Technical talent remains valuable as systems and ambition scale
- 15:07 – 15:30
A workflow to deploy today: daily summaries + chief-of-staff prompts across apps
Thibault suggests concrete high-ROI workflows: daily highlight summaries, staying on top of what people think about your product, and “chief of staff” routines spanning Gmail/Calendar/Docs. The goal is a repeatable morning brief that prepares you for decisions.
- •Set up daily summaries (product sentiment, highlights, unanswered threads)
- •Use multi-tool prompts: Gmail + Calendar + Docs for a daily brief
- •Make the agent ‘chief of staff’ with tailored prep requirements
- •Use agents for brainstorming streamlining and hiring/automation decisions
- 15:30 – 18:47
Live demo: agent runs the inbox, drafts replies, and plans a trip from calendar
They watch the agent pull relevant email threads and calendar context, then propose actions like drafting replies and planning travel around availability. Thibault highlights parallel “agentic threads,” in-app browser use, and the ability to generate artifacts like Google Slides.
- •Agent searches inbox for relevant threads and drafts responses
- •Calendar access enables scheduling and trip planning suggestions
- •Parallel tasks run simultaneously in separate threads
- •Agent can create deliverables (e.g., Google Slides) via in-app browser
- 18:47 – 22:47
Fixing a vibe-coded app: dictation via Speech-to-Text API (and why engineers still matter)
Marina shows an app Codex built for content repurposing, but it lacks voice dictation. Thibault demonstrates turning it into an actionable task by specifying Speech-to-Text API integration, illustrating where technical knowledge (API keys, docs) still unlocks progress.
- •Agents can overbuild (‘too fancy’) without tight constraints
- •Translate a feature need into an implementation plan (STT API integration)
- •Technical gaps: API keys, documentation, and wiring services together
- •Expect mainstream tooling to remove some setup overhead soon
- 22:47 – 24:54
What Thibault uses Codex for: notes, strategy, memory, and unanswered messages
Thibault shares his personal stack: heavy coding use, strategic planning/narrative work, and replacing a separate notes app by writing directly into Codex to build memory. He also uses it to track unanswered important emails/Slack messages.
- •Codex as a thought partner for strategy and narrative sequencing
- •Notes captured directly in Codex to accumulate memory/context
- •End-of-day export: prompting to dump notes into a document
- •Tracking important unanswered communications across channels
- 24:54 – 25:52
Live demo: pulling LinkedIn analytics + turning workflows into reusable skills
They run a computer-use agent to navigate LinkedIn, export analytics, and save data into a spreadsheet—then iterate to request more granular metrics (impressions per post). Thibault explains converting repeated workflows into “skills” that can be run on demand or scheduled.
- •Computer-use agent navigates websites and exports analytics files
- •Iteration improves outputs (from a single export to per-post metrics)
- •Voice-driven operation speeds up commanding the agent
- •Skill creator productizes a workflow into a reusable automation
- 25:52 – 27:45
Career & education: dropping the PhD, following energy, and building in fast cycles
While the agent runs, Marina asks about Thibault quitting his PhD after two weeks and whether that advice holds today. Thibault explains he didn’t want to commit years to one topic, preferred startups, and repeatedly followed instincts toward environments with rapid progress.
- •He left the PhD due to misfit with long commitment to a narrow topic
- •Startups matched his desire to try many ideas quickly
- •Career path: startup → Google → DeepMind → OpenAI
- •Principle: follow instincts and pursue work that gives energy
- 27:45 – 31:07
The ‘personal tailor’ model of AI: authentic conversation replaces prompting as the key skill
Thibault forecasts dramatic life changes in 3–5 years, with benefits arriving even if you don’t actively engage. AI becomes like a tailor who ‘gets you’—the differentiator shifts from promptcraft to asking real questions and showing up authentically in conversation.
- •Future benefits won’t depend on prompt skill or constant engagement
- •AI becomes personalized, context-aware, and proactively supportive
- •Key skill: asking the right questions and engaging authentically
- •Ambient intelligence supports individuals and society in the background
