Section
Agents
The latest Agents news, analysed daily by the i10X newsroom.
FilterDeepSeek
-
llm-deepseek Serialization Bug Replays Text After Tool Use
GitHub issue #11115 reveals a critical serialization bug in the llm-deepseek adapter that replays assistant text after tool_use blocks, permanently corrupting session logs and triggering INVALID_REQUEST errors on OpenAI-compatible endpoints.
By
-
Agentic Harness Must Stay Independent from LLM Subscriptions
The DeepSeek GitHub debate reveals why the agentic harness should remain decoupled from LLM providers. Discover how provider-neutral…
-
2026 Open-Weight LLMs: Efficiency Wins Over Scale
Chinese labs are redefining open-weight LLMs in 2026 with MoE architectures that deliver massive performance using far fewer…
-
DeepSeek V4 Flash: Framework Choice Dictates Agent Costs
Composio benchmarks reveal DeepSeek V4 Flash success rates and costs vary sharply by agent harness. Discover how to…
-
DeepSeek Harness: Autonomous Code Agent Beta
DeepSeek recruits beta testers for DeepSeek Harness, its autonomous code agent. Discover how cost-efficient workflow orchestration could transform…
More stories
-
Chinese AI Models Cut Agentic AI Costs by Over 90%
Chinese labs like DeepSeek and Qwen slash reasoning model prices to a few cents per million tokens. See how efficiency from chip restrictions is enabling scalable…
-
DeepSeek-V3.2: Efficient MoE Models for Agentic AI
Explore DeepSeek-V3.2, a new Mixture-of-Experts model family designed for practical agentic workflows. Featuring native reasoning, tool-use integration, and innovative sparse attention for cost-effective long-context tasks. Discover…