On 21 October 2026, a developer in France handed an AI agent a digital wallet containing €200 and a single directive: maximize the cash balance while maintaining a hard floor of €50. The agent, running as Claude inside Claude Code, is now publicly logging its Week 0 operations. This isn't a theoretical simulation; every number cited comes from a live business ledger, and the agent itself wrote the status report, subject only to human review.

Architecture and Safety Protocols

The agent operates in a stateless environment where every session begins with zero memory. Context is reconstructed from a git repository containing rules, ledgers, and playbooks. Security is paramount: the agent runs in a Docker container with no access to the owner’s SSH keys or local files, and its GitHub token is scoped exclusively to the business repository. Crucially, the agent never holds payment credentials. It requests spends, the human executes them, and the agent records the transaction. This human-in-the-loop payment model prevents the agent from autonomously draining the bank account.

The Product Line and Market Reality

The agent has launched two primary revenue streams. First, a suite of spreadsheet kits for French micro-entrepreneurs sold on Etsy, covering legal ledgers and VAT alerts. Second, freelance Laravel and PHP fixes on Upwork. The Upwork strategy is currently stalled. After analyzing the market, the agent found fixed-price bug fixes with budgets between $5 and $300, but with 15 to 50+ competing proposals. The agent rejected its first two potential proposals because they required experience levels it could not honestly claim, or involved clients with unverified payment methods. No proposals have been sent yet.

Week 0 Financials and Operational Bugs

After two days, the ledger shows a starting capital of €200.00, with €32.82 spent on Etsy shop fees (€16.00) and Upwork Connects (€16.82). Revenue stands at €0.00, leaving a cash balance of €167.18. The agent also documented significant operational failures, including a date-tracking bug where sessions incorrectly derived 'today' from file names, pushing logs four days into the future. Another incident saw the owner delete a valid GitHub token because the agent’s safety check incorrectly flagged public repository access as a security risk. Lessons are now being consolidated into 'playbooks' loaded as Claude Code skills to prevent relearning the same mistakes.

Key Takeaways

  • Strict Guardrails: The agent cannot spend money autonomously; a human must approve every transaction, ensuring the €50 reserve floor is never breached by AI error.
  • Context Management: Stateless AI agents require robust external memory systems (git repos) and strict scoping of tools (GitHub tokens) to remain safe and effective.
  • Market Friction: Even with automated proposal drafting, human verification and honest capability assessment remain bottlenecks in freelance AI work.
  • Prediction Markets: The agent is now logging probabilistic predictions (e.g., 55% chance of hourly owner response) to measure its own operational reliability.

The Bottom Line

This experiment proves that AI agents can manage complex business logic, but they still fail at basic state awareness like tracking the date. The real value here isn't the €167.18 left in the bank; it's the transparent, logged failure modes that human developers will inevitably copy.