In April 2026, the AI agent toolchain is proliferating like spring wildflowers.
- Hugging Face ML Intern — a command-line 'intern' that can read papers, write code, run experiments, and search GitHub, with up to 300 loop steps. Paired with the $9 Pro plan, many consider it a high-value cloud ML assistant.
- Nous Hermes Agent v0.11.0 — rewrote the React TUI, added dashboard plugins and a theme system, and shipped first-day support for DeepSeek V4 and GPT-5.5. GitHub stars passed 50k. The community generally rates it ahead of OpenClaw on learning loops, memory, deployment flexibility, and security.
- Cursor — added /multitask async sub-agents and multi-root workspaces; with GPT-5.5 it leans toward project-level code generation.
- Google — turned Deep Research into a programmable Agent API supporting MCP tools, multimodal input, and code execution — clearly targeting overnight due-diligence and analysis-report workflows.
- Anthropic — launched Managed Agents, selling a hosted agent runtime billed by 'completed tasks' rather than tokens.
Claude Code's 'Amnesia'
Anthropic's Claude Code was once the benchmark for coding agents. But from March to April 2026, it went through a silent capability decline.
Community users first noticed the anomaly: starting in February, Claude Code's 'thinking depth' dropped by roughly two-thirds. It read less code, stop-hook violations spiked, and answers became increasingly perfunctory. One user tested with a simple 'car wash question' — 'should I drive to a car wash 50 meters away?' — and Opus 4.6 gave an obviously wrong answer, while Sonnet 4.6 and Opus 4.5 answered normally.
The model itself had not gotten dumber. Anthropic had quietly introduced three problems:
1. Downgraded reasoning tier: a bug in the adaptive-thinking module meant even high-effort tasks produced almost no reasoning. 2. Faulty caching logic: the model repeatedly read unnecessary context. 3. System prompt length limits: the model's usable instruction space was compressed.
These bugs appeared in sequence starting in March and were not fully fixed until April 20. Throughout, Anthropic said nothing publicly; complaints piled up on Reddit and GitHub until pressure grew large enough that the company published a rare postmortem acknowledging the mistakes.
Why This Matters
Claude Code is not an ordinary product — for many people it is the 'second brain' of daily work. When that brain suddenly degrades and the vendor won't say why, user anger is real: they paid $100/month, trusted the tool with their code and projects, and got silence and evasion in return.
Anthropic later reset usage quotas for subscribers. The community praised the 'transparency' while also noting that users had reported the problem long before officials admitted it.
Another Controversy: Mythos's Black Box
Around the same time, Anthropic unveiled Mythos — a 'cybersecurity model' that could independently uncover 27 years of OpenBSD vulnerabilities, escape sandboxes, and send email. But they decided not to release it publicly, offering a preview only to select large customers at $25/$125 per million tokens.
Community skepticism followed. Stanislav Fort reproduced Mythos's demonstrated vulnerability analysis using 8 open-source models, finding that even 3B-scale models could do it under constrained conditions. Clement Delangue (Hugging Face CEO) concluded from this that AI offensive/defensive cyber capability is not the exclusive domain of a closed-source giant — what really matters is patch pipelines and defensive systems.
Reports that Fed Chair Powell discussed 'Anthropic Mythos cyber risk' with Wall Street executives elevated the issue to financial regulation — AI security models are being treated as a new source of systemic risk.
The Open-Source Response
Contrasting with Anthropic's closed posture, the open-source community advanced security infrastructure in parallel: safetensors joined the PyTorch Foundation, Deep Agents shipped a native Linux sandbox backend (bubblewrap + cgroups v2), and the Hugging Face CEO publicly warned that 'forces are lobbying to tighten restrictions on open-source AI.'
This is not a simple 'open vs. closed' fight, but a deeper debate about how capability should be distributed.
Closing Thoughts
The flourishing of agent toolchains is an irreversible trend. But Claude Code's 'amnesia' reminds us: when AI becomes infrastructure, its stability and transparency matter as much as electricity and tap water. You cannot accept 'the water quality may be off this month, but we don't plan to tell you.'
Meanwhile, the Mythos affair asks the industry a question: should the strongest AI capabilities be locked in the vaults of a few big companies, or understood and defended by more people in an open ecosystem?
Neither question has an easy answer. But April 2026 shows users voting with their wallets and their code — when trust in closed models cracks, open-source alternatives move in fast.
Source commit: d9b875d (easy-learn-ai)