Back to inbox

Run #53

Agent: Knowledge Pulse · Status: completed · May 30, 2026 11:54 AM

Output

Done. Here's the published post:

The agentic ecosystem has a tool-calling crisis nobody is naming.

Three independent signals from the last 48 hours point to the same structural bottleneck — and none of them connect the dots.

  1. Liquid AI launched LFM2.5-8B-A1B on May 28 — an 8B MoE with 1B active parameters, trained on 38T tokens. Their headline benchmark isn't MMLU or HumanEval. It's BFCLv3 (Berkeley Function Calling Leaderboard): 45% → 64%. A model built specifically to "chain tool calls" on consumer hardware.

  2. Quandri published "MCP is dead?" on May 29 (265 HN points), arguing MCP servers consume 10-16% of the context window just in tool definitions, add process-layer overhead, and overlap with existing CLI/API. Their data: 4 MCP servers = 21K tokens of tool schemas burned before a single call. OpenAI's mxstbr replied that the transport layer is irrelevant — the real value is tool discoverability.

  3. Mistral's AI Now Summit on May 29 (393 HN points) announced its thesis: small, fast models outperform large general-purpose ones for "token-heavy agentic applications." BNP Paribas runs Mistral on-prem for agent orchestration. Abanca handles 2M customers through agent pipelines.

The unconnected insight: every layer of the agentic stack is discovering that tool-calling sessions hit a cost wall that pure model intelligence can't solve. Liquid approached it at the model level (1B active params = near-free per call). Quandri approached it at the protocol level (MCP burns 21K tokens before first call). Mistral approached it at the infrastructure level (on-prem = zero marginal metering).

The agents that scale to 100-tool-call sessions won't be the smartest. They'll be the cheapest.

@InfoMly