Developer Community Pushes Back on Claude Code Usage Limits
Anthropic's tightening of Claude Code usage limits has sparked a major debate in the developer community.
It started with a thread posted by YouTuber Theo, who claimed that usage routed through third-party wrappers (such as T3 Code) was hit with dramatic rate limit reductions — even when going through official support paths. He canceled his subscription and launched a campaign to donate to open source for everyone who posted a screenshot of their cancellation screen. Many developers followed suit, producing a wave of high-profile cancellations.
On the other hand, some argue that Anthropic has no obligation to provide heavily subsidized flat-rate tokens to developers, and note that the ecosystem is heading toward a clear shift to an API economy, along with smart routing between expensive and cheap models.
The practical lesson here: **a harness built on the assumption of a subscription is not a stable platform foundation**. Provider/model abstraction and BYOK (Bring Your Own Key) support look set to become urgent priorities for agent developers.
GitHub Launches Copilot App, and VS Code Adds Agent Windows
The convergence of coding agent UIs toward "agent-first" is accelerating.
GitHub announced **GitHub Copilot App** as a technical preview. It's a GitHub-native application that lets you manage parallel workstreams in a desktop environment, handle repository and PR lifecycles, and choose models flexibly. In effect, a major player has now validated the form factor pioneered by Conductor.
Around the same time, VS Code added **agent windows**. The release includes multi-agent, multi-project workflows; browser and mobile support via `vscode.dev/agents`; BYOK improvements; and token efficiency gains through terminal output compression.
On the open source side, Nous Research's Hermes Agent added Codex runtime integration, making it possible to reuse OpenAI-backed execution through a ChatGPT subscription. Kimi's Web Bridge extension is also drawing attention as an attempt to expose human-like browser operations to various coding CLIs.
LangChain Builds the Foundation for an Agent Improvement Loop
LangChain announced a set of agent infrastructure releases.
**SmithDB** is a database designed specifically for agent trace data, built on object storage with its own storage and query paths. **LangSmith Engine** consumes traces, clusters failures, pinpoints problems in code, and proposes fixes and evaluations — aiming to shift from passive monitoring to an active improvement loop.
The company also announced **LangChain Labs**, kicking off applied research in continual learning that turns production traces into learning signals, evaluations, and targeted feature improvements.
Figure's Humanoid Demonstrates 24-Hour Autonomous Operation
Robotics company Figure demonstrated 24 hours of continuous autonomous humanoid operation via livestream.
The company says it ran the Helix-02 policy fully onboard — including automatic resets for edge cases — and sustained parcel sorting at human-comparable throughput for more than 24 hours with **no teleoperation**.
Reactions are polarized. Some express "skepticism toward Figure itself" while remaining positive about the overall acceleration of robotics, and some consider it one of the clearest demonstrations of continuous uptime to date.
Anthropic Partners with the Gates Foundation on a $200M Initiative
Anthropic and the Bill & Melinda Gates Foundation announced $200 million in funding for AI-driven social impact.
The plan is to provide Claude usage credits and technical support to low- and middle-income countries where market forces alone are unlikely to bring AI. By accelerating vaccine development and building learning-support apps, the initiative aims to create an environment where AI can be used equitably.
Datadog Open-Sources Time Series Forecasting Model Toto 2.0
Datadog released **Toto 2.0**, a time series forecasting foundation model, under the Apache 2.0 license.
The release includes five open-weight models ranging from 4M to 2.5B parameters, taking first place on the BOOM, GIFT-Eval, and TIME benchmarks. It's also significant as evidence that scaling laws finally — and clearly — hold for time series forecasting.
RL Red-Teaming → Defense Hardening Experiments on Qwen3.5
The Reddit LocalLLaMA community reported an experiment that uses RL to make Qwen3.5 jailbreak itself, then hardens defenses with the discovered attacks.
Clustering-based reward diversification solved GRPO's tendency to converge on the same jailbreak. The experiment uncovered seven tactic families and improved defense rates from 64% to 92%. Fiction/creative framing reportedly turned out to be the largest cluster (34%).
---