Step 1 — paste your prompt versions

Open AgentPrompt and paste prompt version A and version B. They can be system prompts, task prompts, or any instruction text. AgentPrompt compares them without you re-reading both in full.

Example:

A: "You are a support agent. Always cite a KB article." B: "You are a support agent. Cite a KB article and offer a refund under $20."

Step 2 — pick what to show

Choose Diff (the semantic delta — what behavior changed), Regression (did B drop a guardrail A had?), or Both. The regression check is what tells you the edit weakened the agent, not just that it changed it.

Step 3 — review and gate

Read the behavioral diff and the regression result. If a guardrail was lost, fix version B before it ships — the same discipline as a failing unit test. Attach the result to the change so a teammate can review in seconds.

Step 4 — keep the history

Save the version, roll back a bad edit with one click, and export the history (version + diff + regression result + timestamp) for audit. Re-run on every release, because drift is the default.

An honest limit

AgentPrompt reduces obvious failure modes and makes behavior reproducible. It does not guarantee the agent is safe or compliant, and it is not a legal opinion. Conformity stays the deployer's responsibility across the full system.

Authoritative references

  • EU AI Act (Reg. 2024/1689): https://eur-lex.europa.eu/eli/reg/2024/1689/oj
  • AI Act explorer: https://artificialintelligenceact.eu/
  • OWASP LLM Top 10 (2025): https://owasp.org/www-project-top-10-for-large-language-model-applications/