Coding agents

Anthropic makes the case for a single agent, Nanbeige4.3-33B, Cline Desktop 0.0.49: the October 26 roundup

Anthropic backs a single agent over subagents, Nanbeige ships an open 33B model, Cline Desktop shows diffs and late data centers now cost $5M a day.

By Audryk · October 26, 2026

• Coding agents
• Models
• Cloud & infrastructure
• #Anthropic
• #agents
• #open source
• #Cline
• #data centers
• #Hugging Face

This roundup covers the weekend of October 24 and 25, 2026 and the night of the 26th, with five stories our earlier editions had not covered: an architecture guide from Anthropic, an open model for coding agents, two Cline releases, delay penalties at AI data centers and a funding round in generated video. Unless stated otherwise, figures come from official sources.

Anthropic: fewer subagents, one agent with skills

Anthropic has published a guide titled "When to fold subagents into your main agent loop" (Anthropic article), covered on October 25 by The Decoder. Its Applied AI team describes what it sees at enterprise customers.

• The finding: across several deployments it compared, a single agent with skills often beat both one giant prompt and a subagent architecture on quality, and often at lower cost and latency per task. One customer measured a roughly 2x faster run with a single agent loop.
• The method: fold subagents in one at a time, turning their prompts into skills and their API calls into MCP tools, and compare an end-to-end eval before and after. Anthropic notes that past 30 to 50 tool definitions in the system prompt, tool choice tends to degrade.
• The exceptions: subagents still help when a single agent's context measurably degrades, for parallel tasks with throwaway context, or when the agent struggles to pick its tools.
• A kit: a reference repository, single-agent-loop-kit (Apache 2.0 license), guides the migration. Anthropic also says it cut the system prompt of Claude Code by more than 80% for its recent models.

Nanbeige4.3-33B: an open 33-billion-parameter model for agents

Nanbeige, the lab inside Chinese recruiting platform BOSS Zhipin, put Nanbeige4.3-33B and its base model on Hugging Face on October 24 under the Apache 2.0 license (model card, MarkTechPost).

• The model: 33 billion parameters, dense (no mixture of experts), text only in English and Chinese, with a 524,288-token context window. Nanbeige presents it as a "teacher" model meant to distill a future Nanbeige4.3-3B.
• Reported scores: against Qwen3.8-27B, 80.9 vs 73.0 on Terminal-Bench 2.1 and 65.5 vs 42.5 on DeepSWE; it trails on SWE-Bench Pro (60.3 vs 61.7). These are the vendor's numbers, measured with its own harnesses.
• Running it: the card gives Transformers, vLLM and SGLang recipes; the server examples split the model across 4 GPUs. BF16 weights weigh about 66 GB. The model joins our model comparison.

Cline Desktop 0.0.49 and Cline CLI 3.0.74 (Oct 26)

Cline shipped several releases overnight (Cline Desktop 0.0.49 notes, CLI 3.0.74 notes).

• See what the agent changed: a "Changes" button lists the files changed in the last turn or across the whole session, with a diff of each. A project files panel and an integrated terminal now sit beside the conversation.
• Local models: a new settings page manages Ollama and LM Studio models, with setup guidance and the memory used by the loaded model (documentation).
• Reliability: opening a task already running in another Cline app now attaches to it instead of starting a second one, and model requests retry at most 3 times instead of 5. With the gemini provider, Cline no longer sends native Google Search, which made every turn fail on free-tier keys with a 429 quota error.

AI data centers: more than $5 million per day of delay

According to The Information, as relayed on October 25 by WebProNews and Benzinga, tenants of new AI data centers now ask for more than $5 million in penalties for each day that power or buildings are delivered late. In the spring, broker Marsh put together "delay-in-delivery" insurance for a developer whose tenant wanted $1 million a day. This is a press investigation, not figures published by the companies. In its own analysis, Marsh notes that standard delay-in-start-up cover generally only responds to delays caused by physical damage (Marsh).

Higgsfield raises from NVIDIA and TCV (Oct 23)

AI video and image platform Higgsfield announced an extension of its Series B with TCV and NVIDIA coming in, without disclosing the amount (Higgsfield announcement). The company reports more than 40 million users, an annualized revenue run rate above $1 billion reached in September, 18 months after launch, and 12 trillion tokens processed per month. These figures are reported by Higgsfield. The original Series B, led by DST Global in August, raised $400 million at a $5.4 billion valuation, according to Bloomberg.

The number: $5 million per day

That is the delay penalty some AI data center tenants now ask for in new leases, according to The Information, up from $1 million in the spring. Power is becoming the main risk on these projects.

What it changes

On the software side, the trend is simplification: one agent, skills loaded on demand, and tools like Cline that clearly show what the agent changed. On the hardware side, demand is such that every day a data center runs late is counted in millions.

Sources

• Anthropic: Building agents, when to fold subagents into your main agent loop (verified 2026-10-26)
• Anthropic: single-agent-loop-kit repository (verified 2026-10-26)
• The Decoder: Anthropic says one AI agent beats many subagents (verified 2026-10-26)
• Hugging Face: Nanbeige/Nanbeige4.3-33B (verified 2026-10-26)
• MarkTechPost: Nanbeige Releases Nanbeige4.3-33B (verified 2026-10-26)
• Cline: Desktop 0.0.49 release notes (verified 2026-10-26)
• Cline: CLI 3.0.74 release notes (verified 2026-10-26)
• Cline: Local Models in Cline Desktop (verified 2026-10-26)
• WebProNews: AI Data Center Tenants Now Want $5 Million a Day When Projects Run Late (verified 2026-10-26)
• Benzinga: Data Center Late Fees Hit $5 Million or More Per Day (verified 2026-10-26)
• Marsh: The new age of contractual risk allocation in data center development (verified 2026-10-26)
• Higgsfield: Series B extension with TCV and NVIDIA (verified 2026-10-26)
• Bloomberg: Nvidia, TCV Invest in AI Video Startup Higgsfield (verified 2026-10-26)

Spotted an error in this article? Report it · Our corrections policy