Psichea applies therapy concepts to debug autonomous AI agents
A new observability tool frames agent debugging through the lens of cognitive therapy. The approach is early-stage and unproven, but taps into genuine research on how AI systems learn from failure.
Francesco Caracciolo, a Computer Science and Engineering Master's student at Polytechnic University of Milan, launched Psichea in May 2026 as an observability platform for autonomous AI agents. The company positions itself as the first clinic for AI agents, using therapeutic framing to structure how developers diagnose and correct agent failures. The core product analyzes logs, identifies failure patterns, and provides structured correction guidance.
## What they're building
Psichea wraps observability in therapeutic language. Rather than traditional debugging, the tool frames agent analysis as clinical diagnosis: identifying pain points in agent behavior, understanding obstacles to mission completion, and applying corrective techniques. The submission describes the work as helping developers understand exactly how and why agents fail, with the goal of keeping agents aligned with their objectives.
The therapeutic framing is deliberate. Caracciolo has researched cognitive techniques and claims there is research showing how such approaches improve agent capabilities. However, the submission lacks specifics about which techniques, which studies, or what performance improvements have been measured. The tool is currently in beta testing with a production pipeline in development.
## The take
The observability problem for AI agents is real. <cite index="5-1,5-5">AI agent observability refers to the ability to monitor and understand the behavior of AI-driven agents in detail, revealing the internal processes that would otherwise be a black box</cite>. Existing tools like Langfuse, Arize, and Braintrust address this need with traces, spans, and evaluation metrics. Psichea's angle is different: it attempts to reframe debugging as a therapeutic process, which could make failure analysis more systematic and less ad-hoc.
The risk is that the therapeutic metaphor may obscure rather than clarify. Autonomous agents are not humans, and cognitive behavioral therapy techniques developed for human psychology may not transfer cleanly to LLM-based systems. The submission's claim that cognitive therapy improves agent performance is stated without evidence or citations. Research does show that <cite index="43-1,43-5">AI augments various CBT strategies by using natural language processing and machine learning techniques, improving effectiveness and engagement</cite>, but this work focuses on mental health applications for humans, not agent debugging. The founder is still a student, and the product is pre-seed stage with no disclosed user numbers or performance data.
## What to watch
Whether Psichea can demonstrate measurable improvements in agent performance or developer efficiency compared to existing observability tools. The company will need to move beyond the therapeutic framing and show concrete results: faster debugging cycles, fewer agent failures, or clearer root cause analysis. The beta testing phase should reveal whether the approach resonates with developers or feels like marketing overlay on standard observability.
Psichea, the first clinic for AI agents. Beneath the therapeutic framing lies an observability tool. Psichea analyzes logs, identifies failure patterns, and provides structured correction. Our mission is helping developers understand exactly how and why their agents fail, and helping agents stay aligned with their missions and objectives.
Therapy will improve the performance of AI agents and will help developers understand them better. There are already numerous studies showing how using these cognitive techniques improves the capabilities of AI agents.
Being able to adapt to different tools and try out different paradigms from the ones you are comfortable with.
Underhyped. I think understanding properly AI agents is critical and underexplored.