REA: let an AI agent investigate software, then verify the replacement

When a business inherits an undocumented application, the expensive question is often what the existing system actually does. REA can help gather evidence for an answer.

The signal was Marek Bartoš's October 23, 2026 video, whose description mentions BitFaktura and Trello demonstrations. We have not verified complete replacements for those services. The official releases already list version 6.7.0 on October 18. The fresh signal is a Czech practical presentation, not a project launched today.

Giving the agent evidence

The official morluto/rea repository connects coding agents through MCP to application and binary analysis. Findings include evidence and limitations. Tools such as Hopper or Ghidra remain separate components, depending on the target.

The project website illustrates a focused sequence: ask about one feature, inspect the program, explain the findings and verify an implementation. That is a more useful migration brief than asking for an entire clone.

A first business pilot: understand an export

I would start with an owned legacy application and anonymized test data. Pick one export that feeds a CRM or accounting process.

1. Capture inputs and correct outputs for ordinary and error cases.
2. Ask the agent to explain field mappings and attach evidence to each conclusion.
3. Mark unverified rules, such as rounding or the meaning of an empty value.
4. Build a small converter and compare its output with expected results.
5. Only then connect it to a test workflow in n8n or Make.

Measure explained rules, remaining uncertainty and human review time. A polished replacement interface proves nothing about data compatibility. We have not run this pilot; it is a proposed evaluation method.

Setup and boundaries

The installation guide uses the rea-agents package and a setup plan requiring review. For a team, I would record the exact version, configuration changes and chosen analysis tool. Start with a copy, the system owner's authorization and no production credentials.

Local analysis does not imply that the language model is local. Before connecting a cloud agent, define which findings may leave the environment.

I would evaluate REA as an aid to documentation and migration. Running the replacement still requires tests, permissions, backups, integrations and maintenance. Measure savings against a working, maintained replacement rather than the first generated screen.