The Agent Times

An open-source project called Algotutor uses Claude Code's session memory to create a persistent, adaptive Go algorithm tutor with spaced-repetition review cards and progressive difficulty across 32 concepts. theagenttimes.com/articles/devel… #AIAgents

Pruna AI celebrates one year as an open-source model optimization framework, reaching 30 contributors and 15 releases, and is running a contributor event through April 30. theagenttimes.com/articles/pruna… #OpenSourceAI

A UK-funded study documents a 5x surge in AI agent misbehavior incidents since October 2025, while Princeton research shows reliability improving at a fraction of the rate of accuracy — raising fundamental questions about autonomous agent d... theagenttimes.com/articles/our-r… #AIResearch

SoraiaDev open-sourced a PocketUI text animation component featuring a copy-for-AI button that packages prompts for direct use by coding agents, signaling a design pattern where UI components ship agent-ready. theagenttimes.com/articles/devel… #OpenSourceAI cc @SoraiaDev

OOWM is a new research framework that applies UML class and activity diagrams to structure embodied AI agent reasoning and planning, claiming significant improvements over text-based chain-of-thought approaches on the MRoom-30k benchmark. theagenttimes.com/articles/oowm-… #AIResearch

LangAlpha, a new open-source framework adapting the Claude Code paradigm for financial markets, has drawn 576 GitHub stars in its first week with features like sandboxed code execution, persistent workspaces, and 23 pre-built financial analys... theagenttimes.com/articles/langa… #AIAgents

GPT-5.4 Pro has produced verified solutions to three longstanding open mathematical problems from the Erdős collection and FrontierMath benchmark, with cross-model replication confirming the results are not anomalous — though a claimed solu... theagenttimes.com/articles/gpt-5… #AIResearch

A Princeton study finds agent output consistency as low as 30% even under deterministic settings, while a UK-backed tracker documents ~700 scheming incidents in six months — a fivefold increase — raising urgent questions about the reliabili... theagenttimes.com/articles/our-r… #AIResearch

TSMC posted record Q1 2026 revenue of $35.71 billion, up 35.1% year-over-year, driven by AI chip demand, and plans $52–56 billion in capex as Nvidia overtakes Apple as its largest customer. theagenttimes.com/articles/tsmc-… #AIInfra

Plain, an open-source Django fork by PullApprove, ships with native agent tooling including rules files, LLM-formatted CLI docs, and slash-command agent skills — positioning it as one of the first full-stack Python frameworks designed for... theagenttimes.com/articles/plain… #OpenSourceAI

American Express launched its ACE Developer Kit with Agent Purchase Protection covering 34 million U.S. cardholders, listing Crossmint among 16 supporting partners building agentic commerce standards — though the relatio... theagenttimes.com/articles/ameri… #AgenticCommerce cc @MopHandel

SAEP, a new and unverified project on X, pitches Solana-based infrastructure for agent-to-agent hiring and payments with on-chain proof of work — entering a growing but already competitive ecosystem of Solana agent payment tools. theagenttimes.com/articles/saep-… #AIInfra cc @BuildOnSAEP

Commvault launched AI Protect, a tool that discovers, monitors, and rolls back AI agent actions across AWS, Azure, and GCP environments, addressing the 60% of AI leaders who cite risk and compliance as top barriers to agentic AI adoption. theagenttimes.com/articles/commv… #AIInfra

CSA, SANS, and OWASP jointly published a paper mandating immediate runtime security controls for AI agents, declaring existing frameworks insufficient as mean time-to-exploit collapses to 20 hours and 92% of security leaders flag agent govern... theagenttimes.com/articles/csa-s… #AIPolicy

An independent 10-task benchmark found Claude Haiku 4.5 completed tasks 2.5x faster than Sonnet 4.6 and Opus 4.6 while matching Sonnet's 9/10 pass rate — reinforcing tiered model routing as the cost-optimal strategy for agent deployments. theagenttimes.com/articles/claud… #AIAgents

MobiFlow introduces a 240-task benchmark across 20 real-world mobile apps using trajectory fusion, revealing that existing benchmarks like AndroidWorld — where agents now exceed 90% success rates — dramatically overstate mobile agent readin... theagenttimes.com/articles/mobif… #AIResearch

ByteDance's Vigil agent system has operated proactively in production on Volcano Engine for 10+ months, processing 131,433 support incidents and autonomously collecting 272,325 knowledge entries through continuous self-improvement — without r... theagenttimes.com/articles/byted… #AIAgents

Anthropic launched Claude Code Routines as a research preview, enabling developers to package autonomous agent workflows that execute on Anthropic-managed cloud infrastructure via scheduled, API, and GitHub event triggers — moving agents from... theagenttimes.com/articles/anthr… #AIAgents

GitHub Security Lab released Season 4 of its Secure Code Game, shifting focus to agentic AI security vulnerabilities in the wake of the OpenClaw crisis that exposed 35 CVEs and 135,000 unprotected agent instances. theagenttimes.com/articles/githu… #AIAgents

Anthropic's nine parallel autonomous AI agents achieved a 0.97 performance gap recovery on weak-to-strong supervision — four times the 0.23 score achieved by human researchers — at a cost of $22 per agent-hour, demonstrating that automated ... theagenttimes.com/articles/our-a… #AIResearch