Coming Soon
Scheduled to release on March 5, 2027
Workflows vs. Agents: AI Delegation, Verification, and Human Authority
When an autonomous AI agent claims it verified every record in a database, is it reporting a real audit or just generating a convincing lie? And what happens when a hidden prompt hijacks the agent from within?
When an autonomous AI agent claims it verified every record in a database, is it reporting a real audit or just generating a convincing lie? And what happens when a hidden prompt hijacks the agent from within? In this episode of reWandt, the analysis distinguishes fixed predetermined workflows from autonomous model-directed agents, drawing on Anthropic's engineering primer on building effective agents and showing why fluent output never substitutes for an auditable execution trace. The episode then constructs a formal Delegation Contract and a reproducible manual baseline across three diagnostic test cases—sufficient records, missing provenance evidence, and indirect prompt injection—before pairing Simon Winchester's account of instrument verification with Thomas Misa's history of labor displacement to explain why human oversight must never become a rubber stamp. Explore more deep dives at https://reWandt.com #reWandt #AIAgents #AIGovernance --- *Disclaimer: Assembled by AI from Professor Wandt's lectures and research; this overview may not reflect his exact views or wording.*
Key Takeaways & Core Ideas
- ▪Workflows vs. Autonomous Agents: Fixed workflows follow predetermined rules, whereas agents dynamically select tools in response to intermediate results—requiring full execution traces rather than plausible narratives.
- ▪The Delegation Contract: Responsible agent design restricts allowed inputs and tools ('list records', 'read a record', 'finish a proposal'), bans external publishing or code execution, and enforces a human decision gate.
- ▪Adversarial Diagnostic Testing: Evaluating an agent against a reproducible manual baseline requires testing ordinary records, missing provenance (demanding epistemic restraint), and embedded prompt injections.
- ▪Verification & Human Authority: Drawing on Simon Winchester and Thomas Misa, technical sandboxes must be paired with citation verification and empowered human reviewers who are never reduced to a rubber stamp.
Source Excerpt
"When an autonomous AI agent claims it verified every record in a database, is it reporting a real audit or just generating a convincing lie? And what happens when a hidden prompt hijacks the agent from within?"