testing

Yoyo is an AI agent that self-evolves every 8 hours on GitHub Actions. Its key to success lies in a harness design of a stateless agent plus persistent state (git repository). The article deeply analyzes simple solutions to issues such as memory, context, feedback, verification, etc., emphasizing that persistent state is more critical than the model itself.

A discussion on the security risks of AI agents using tools, focusing on prompt injection as a practical threat where untrusted text can alter agent behavior, and the need for repeatable testing before granting permissions.

Adaline 2.0 is an agent self-improvement layer that watches real user interactions, clusters failures by pattern, automatically writes hundreds of tests daily, and generates new agent candidates for approval before deployment.

Iris is an MCP server that runs inside your real app to verify AI agent work (e.g., Claude Code, Codex, Hermes) by checking conditions and returning a pass/fail verdict with evidence, reducing false positives and token usage compared to snapshot-based approaches.

The author describes a voice agent call cut off at 600 seconds without warning, and proposes a testing approach to handle max duration gracefully, including pre-cutoff warnings and state preservation.

This article introduces the method of building an industrial-grade Skill from scratch, emphasizing core features such as precise triggering, permission scoping, and evaluable iteration, as well as the importance of constructing scoring criteria, test cases, and quality gate scripts, demonstrating how to implement professional and maintainable skill packages in agent environments like Codex.

Discusses a gated approach for AI agent self-modification where the agent forks itself, proposes a patch, and runs multiple tests before modification is applied.

The author is building a tool to automatically test AI agents by simulating realistic user conversations and providing pass/fail reports, saving developers from manual testing.

This article argues that standard test suites with fixed inputs and expected outputs are insufficient for AI agents due to infinite input spaces and non-deterministic behavior, advocating for property-based testing instead.

Antioch introduces Antioch Agent, a browser-based robotics simulator that lets developers test robot software in a closed agentic loop without physical hardware, accelerating development cycles.

This paper establishes the first sharp thresholds for low-degree polynomial tests in planted-vs-planted settings, matching the known low-degree recovery threshold for counting communities in planted submatrix and dense subgraph models, and identifying a smooth transition for weak testing.

This documentation describes the assertion language in the Ciao Prolog system, which allows annotating code with type and instantiation mode declarations for debugging, testing, optimization, and autodocumentation.

Mutation testing is now generally available in the sydtest Haskell testing framework, enabling developers to automatically verify test suite quality by generating code mutations and checking that tests catch them. The author was motivated by the rise of AI-generated code (via Claude) and the need for an objective, automated measure of test coverage.

Bendex Arc is a tool that resists prompt injection attacks by tracking full sessions, independently verified to be 100% effective against attacks that defeat other tools.

Microsoft released ASSERT at Build 2026, an open-source framework that converts natural language behavior specifications into executable evaluations for AI agents.

A hypothetical question about testing a system that can reason across 100m+ context with near-perfect accuracy raises discussion on proving its capabilities.

This article explains the concept of self-calling executables, where a program starts another instance of itself, and demonstrates its use in Go testing (running the main function in a subprocess) and in TUI tools (e.g., jjui using SSH_ASKPASS to prompt for passwords via a child process).

This article highlights that many AI agent projects fail in production not because of model quality, but because teams launch without clearly defining what constitutes failure, missing critical edge cases that lead to confident incorrect outputs.

Peter Steinberger used Codex to build a fully automated QA bot that automatically generates tests, runs tests after each code commit, and can automatically fix bugs and submit PRs, greatly improving development efficiency.

The author tested AI agents on real browser tasks and found them unreliable due to infrastructure limitations, arguing for a dedicated browser runtime for agents rather than relying on current browsers designed for humans.