Building AI Self-Healing Agents with Python - 2026

Learn that an agent is a model, tools, and a loop, with the model deciding tool calls and the loop enabling multi-step actions; the course builds self-healing agents in Python.

Set up python environment with uv, create a self healing agent project, install Entropic and pytest, and prepare an Entropic API key for the first API call in .env file.

Define a tool by building a function, a schema, and a registry, then register it and drive a model to count words in text.

Explore the stop reason concept in self-healing AI agents, distinguishing end turn and tool use, and how max token limits affect tool calls.

Explore how a model returns a list of typed blocks and how tools are invoked to fetch price and market cap data, with final results in the second call.

Build and test your first ai agent loop using a stock price tool with Apple, Microsoft, and Google, then explore typos, unknown tickers, and ambiguous questions to enable self-healing.

Explore six AI agent failure modes—wrong output, stuck loop, tool error, hallucination, timeout, cost blowout—and understand loud vs silent failures, plus detection via sanity checks, schema validation, and LLM critique.

Develop the execution monitor to produce a trace of every step—tool calls, model results, and outputs—capturing timing, results, errors, and a confidence score for detection and learning.

Detect silent failures by building a detector that reads the trace and output to emit failure objects, using sanity rules, structural checks, and an LLM self-critique in order for efficiency.

Build a silent detector class in Python that applies three safety checks—sanity rules, repetition detection, and an LLM critique—to flag failures and hallucinations.

Diagnose failures by classifying them into five root causes—bad input, bad tool, bad prompt, bad model, and bad logic—and route targeted recoveries accordingly.

Define a diagnostician with Cause, Diagnostics, and Diagnostician classes to analyze traces and failure codes, using an LLM to output JSON with cause, evidence, fix, and confidence.

Build and test a Python diagnostician to classify failures, capture traces, evidence, and suggested fixes, and validate prompts and retry logic across three scenarios.

Learn how to use prompt self-modification with delta edits to fix failing system prompts, balancing depth with the one-sentence constraint.

Discover how strategy switching lets ai agents generate diverse alternatives, score them, and pick the best approach, including chunked summarization and novel recovery strategies.

Introduce the recovery confidence score to determine whether a recovered output actually solves the task, using semantic comparison with the original failure and a strict 0–1 llm evaluation.

This lecture introduces a 20-test adversarial suite to probe a self-healing agent's honest failure handling before production. It covers typo variants, tool failures, and prompt conflicts.

Learn to use pytest to build and run unit tests for AI agents, exploring test files like test_calculator, fixtures and parameterized tests, and validating self-healing behavior.

Showcases pytest parametrized typo tests and adversary testing against an AI self-healing agent, capturing failures, evidence, confidence, and recovery actions to correct misinterpreted tickers like AAPL, APPL, MSTF, XYZ.

Explore unit tests for silent failures—invented citations, historical fabrication, and partial result disclosures—to build robust, self-healing AI agents with Python.

Explore recovery correctness testing for self-healing AI agents, including recovery logs, polite and genuine failures, and recovery score thresholds. Learn how unit tests across categories validate robust recovery behavior.

Expose a self-healing agent via a Python FastAPI server with uvicorn. Create endpoints like run and health, using logging for production observability.

Explore building self-healing AI agents in Python that know when they fail in production. Share your work with the community to learn, contribute, and ship honest systems that reveal failures.