Failed Runs Dominated Inference Spend in One Harness Eval
A benchmark reports that 64% of inference spend went to runs that didn't pass, across 27,280 runs and 83B tokens. The page is titled "Introducing...

Created by weiqun zou
Latest news, benchmarks, and use‑case guides for AI coding assistants and autonomous agents
Explore the latest content tracked by AI Coding Tools Digest
A benchmark reports that 64% of inference spend went to runs that didn't pass, across 27,280 runs and 83B tokens. The page is titled "Introducing...
The report describes a 320-billion-parameter mixture-of-experts model with dedicated encoders for molecular graphs and biological sequences, feeding a...
Start by separating the two billing camps. The GLM Coding Plan and Kimi Code are quota-based subscriptions, while the DeepSeek API bills per token....
Designing around AI shifts the focus from individual coding speed to how work moves between stages. Four changes stand out:
The post frames SWE-CoWork as a test of whether coding agents can be dependable digital engineers on long-horizon tasks. To pass, agents must gather...
Permission layers usually see only the path text an agent requests, but the kernel follows symlinks until it reaches the actual file. A committed...
The Jev MCP server is a Model Context Protocol implementation that links assistants like Claude and Cursor to your Jev account, giving them structured...
Sessionpipe, announced as open source by @micmakarov, says it connects coding-agent sessions "on any harness" and lets developers save, search,...
Kiro's spec-driven flow turns one prompt into requirements, design, and tasks you review before any code is written. The article argues that catching...
The core dispute is whether agent-run checks justify their token cost. One commenter on the Claude Code team's verification post calls some practices...
Two very different threats are emerging. SPILLWAY, a new arXiv preprint, targets hidden instructions in web pages or documents that could redirect...
The paper introduces Accelerate Bench, with 32 CPU tasks drawn from scientific applications and supercomputing competitions. Agents get a working...
Speed and price look strong on paper, but the quality case is still vendor-reported.
The guide makes a cost case, since agentic workflows generate millions of tokens and DeepSeek V4.1 Flash lists at $0.098 per million input tokens...
Anthropic told IPO investors that Mythos Preview did the work in about three days with "basically no human intervention". The details narrow the...
Long AI-generated diffs can look convincing in review, so check them line by line on sensitive paths.
A repost of a HuggingPapers post describes MIRIX-2.0 as a multi-agent memory system for LLM agents that combines factual and procedural memory. The...
A Show HN post introduces Pulsar as cross-harness agent-to-agent communication and messaging. The author, contrahax, shared it with 15 points and...
Together notes that coding agents resend the whole conversation on every turn. The post argues that prefix caching keeps long sessions fast, but only...