🤖 Agents Weekly: DeepSeek-V3.1, AGENTS.md, URL Context, Context Engineering Tips, Qwen-Image-Edit

DeepSeek-V3.1, AGENTS.md, URL Context, Context Engineering Tips, Qwen-Image-Edit

In today’s issue:

• DeepSeek releases DeepSeek-V3.1

• Qwen has released Qwen-Image-Edit

• OPPO introduces Agent Foundation Models

• Firecrawl releases Firecrawl /v2 with improved search for AI agents

• A multimodal agent with long-term memory

• Context engineering and prompt engineering tips from Anthropic

• AGENTS.md aims to standardize context injection for agents

• Google announced URL Context for the Gemini API

• Top AI devs news, papers, and product updates.

---

---

Top Stories

DeepSeek-V3.1

DeepSeek introduced V3.1, an updated reasoning model designed for the coming agent era. It blends fast non-thinking responses with deliberate reasoning, while improving tool use and efficiency.

• Hybrid inference – V3.1 supports both Think and Non-Think modes, switchable via a new DeepThink button.

• Faster reasoning – The V3.1-Think variant reaches answers faster than the previous DeepSeek-R1-0528 model.

• Agent skills boost – Post-training enhances multi-step reasoning, tool usage, and performance on agent benchmarks like SWE and Terminal-Bench.

• Model update – Built on V3 with continued pretraining on 840B tokens for long-context extension (128K). New tokenizer and chat template included; open-source weights released.

• API upgrades – Clear split between deepseek-chat (non-thinking) and deepseek-reasoner (thinking), Anthropic API compatibility, and strict function calling support.

• Efficiency gains – Stronger performance on complex search tasks with significant improvements in thinking efficiency.

Model | Pricing