40 millions de vues en 12h : Claude "Computer Use" enterre le travail de bureau ?

Core concept: machine control by autonomous agents

Machine control is the stage where autonomous agents manipulate real desktop applications by driving the mouse and keyboard (like a human). This enables automation of legacy and internal software that lacks APIs, extending beyond what chatbots, copilots, and orchestrators could automate.

Evolution (four stages)

1. Chatbots — answer questions.
2. Copilots — in-tool assistance.
3. Orchestrators — coordinate data between apps.
4. Machine control — agents operate desktop apps directly.

Product / feature breakdown and examples

• Claude (Anthropic)
  • Demoed “computer use”: an agent running on a Mac that reads files, navigates apps, takes screenshots, and asks permission before actions.
  • Supports “dispatch”: remote triggering/controlling from a phone.
• Cowork / Cloud Cowork (Anthropic/Claude)
  • Acts like a junior operational assistant on the laptop.
  • Supports scheduled agents, recurring tasks, and cross-app workflows.
• Dispatch
  • Remote control functionality to trigger desktop agents from mobile devices.
• Cloud Code
  • Programmable agents: configure once for scheduled, repeatable execution.
• Perplexity
  • “Cloud manager” approach: uses a dedicated Mac Mini plus cloud orchestration of many models.
  • Offers governance, logs, and permissions, but routes data through cloud servers.
• OS-integrated agents
  • Examples: Google (Gemini), Microsoft (Copilot).
  • Provide OS-level integration and fine-grained permissions, but increase platform lock‑in.
• Open local / open-source agents (“Open Clow”)
  • Favor local control and sovereignty.
  • Higher security risk due to potentially malicious plugins and remote-execution vulnerabilities.

Trust architectures — three competing models

1. Open local
   • Install on your machine, self-host API keys.
   • Pros: maximal theoretical sovereignty.
   • Cons: governance and security problems, hostile plugins.
2. Cloud manager
   • Cloud-run on dedicated hardware with central governance and logs.
   • Pros: better auditability and control.
   • Cons: requires sending data through third-party servers; can be costly.
3. OS integration
   • Embedded by platform vendors.
   • Pros: seamless UX and OS-level permissioning.
   • Cons: vendor lock‑in and data/process funneling to the vendor.

Security, privacy, and limitations

• Data sovereignty trade-offs
  • “Local” execution can still transmit screenshots or data to vendor servers (e.g., Claude sending screenshots to Anthropic’s cloud).
  • Sensitive/legal/compliance constraints make adoption risky today.
• Reliability gaps vs. demos
  • Benchmarks show large gaps between demos and reliable end-to-end results.
  • Example results: one test produced billable deliverables for only ~2.5% of tasks; top models fail >75% on complex office tasks.
  • Machine control is still slower and less reliable than native integrations.
• Product limitations / warnings
  • Anthropic’s prototype: not for legal, medical, or sensitive documents; requires Mac to remain awake; enterprise plans excluded from the prototype.

Practical implications and market analysis

• Immediate productivity winners
  • Professionals with repetitive extraction/compilation/reporting workflows: analysts, product managers, consultants, operations.
  • Repetitive 45‑minute tasks can become a single instruction.
• Shift in value
  • Governance (who delegates, audits, and manages agent autonomy) becomes the main source of financial value, rather than the individual doing manual work.
• Implicit data capture
  • Agents record micro‑behaviors (tab order, hesitations, deleted drafts), creating a “4th layer” of enterprise data: informal processes, shortcuts, and bottlenecks that can be monetized or used to optimize workflows.
• Startup risk
  • Rapid platform integration by major vendors compresses startup windows; viral features can be absorbed by large platforms quickly.

Signals to watch (indicators of mass adoption)

• End of Apple exclusivity — desktop control arriving on Windows (Windows ≈ 70% installed base).
• Availability of enterprise-grade mode — infallible audit logs, per-app granular control, and strict compliance.
• Expansion to phone control — agents operating on iOS/Android and handling calls/forms.

Recommendations and hands-on guidance

• Be skeptical of spectacular demos (Minecraft, flashy one-click clones) — entertaining but not proof of reliable office automation.
• Calibrate risk before deployment
  • Decide scope, supervision level, and pace of autonomy.
• If you have prototype access
  • Try Cloud Cowork on a non‑sensitive project (for example: extract quarterly figures from an email and update an Excel table) to observe workflow differences without manual copy/paste.
• Avoid ungoverned local setups
  • Don’t run unmanaged dedicated hardware (e.g., Mac Minis) if better integrated and governed options are imminent.

Resources, reviews, guides, and tutorials mentioned

• Upcoming comparative matrix (Anthropic, OpenClow, Perplexity, Google, Microsoft) — promised on Patreon.
• Three structured courses promoted:
  1. Foundations of AI.
  2. First Cloud Code project (practical).
  3. Cloud Cowork mastery — includes audio, quizzes, Discord discussion, and direct troubleshooting access.

Benchmarks and reports cited

• Independent autonomous-agent benchmarks report very low end-to-end success rates:
  • Example: 2.5% billable deliverable rate in one 240-order test.
  • Example: >75% failure on complex tasks in another benchmark.
• Forbes report: Claude sessions reportedly jumped ~1500% in March (a user migration metric).

Main speakers / sources referenced

• Anthropic / Claude (primary product discussed)
• Perplexity (cloud manager approach; Computer & Dispatch features)
• “Open Clow” (open/local agent community/project reference)
• Google (Gemini) and Microsoft (Copilot) — OS-integrated approaches
• Terms and names referenced: Entropic / Tropic / Tropique (likely tied to Anthropic’s prototypes or channels)
• Independent benchmarks (unnamed) and the “Apex” benchmark
• Forbes (user adoption statistic)
• The presenter/channel (referred to as “Lia” / the channel author) — provided analysis, courses, Patreon deliverables, and calls to action