Cowerx

About · RSS · Atom · JSON Feed

The journal /

daily.md 2026-09-28

Selected AI news for builders. Every headline links to its source.

What is hot right now, rebuilt every 15 minutes: The Daily Markdown front page.

Start reading this issue

Analyzing and Mitigating Cost-Inefficient Behaviors in Coding Agents

Researchers analyzed 1,200 trajectories from Claude Code and Mini-SWE-Agent to identify three cost-inefficient behaviors that affect 79.00% to 98.00% of coding tasks. They found that developer-designed skills reduce cost by up to 41.73%, which is roughly twice the maximum gain from agent-synthesized skills. Developer-designed skills reduce coding agent costs by up to 41.73%.

Summary by a local model.

Source: arxiv-ai · 2026-09-28

Cartograph: Federated Tool Discovery with Operator-Attested Retrieval for AI Agents

Researchers presented Cartograph, a federated proxy that reduces AI agent tool discovery from full catalog traversal to progressive disclosure. On a 374-tool deployment, the system exposes three proxy tools and uses 475 tokens for a top-5 discovery exchange instead of 42,450. Builders care because the proxy reduces token overhead from 42,450 to 475 for tool discovery.

Summary by a local model.

Source: arxiv-cl · 2026-09-28

Probing Stability-Plasticity Tradeoffs in Agent Memory through Cognitive Experimental Paradigms

Jiaqi Ding and Guorong Wu introduced MemProbe, a framework that uses four cognitive paradigms to diagnose stability-plasticity tradeoffs in agent memory. The authors evaluated six incremental memory systems using a 56-episode diagnostic suite and found that systems with similar aggregate scores exhibit distinct behavioral profiles. Builders care because the framework reveals how systems update, preserve, and organize information beyond final-answer accuracy.

Summary by a local model.

Source: arxiv-cl · 2026-09-28

Not All Memories Are Equal: Hierarchical Collaborative Memory for Validity-Aware Retrieval in LLM Agents

Yufei Shi and five other authors proposed HiCoMER, a framework that manages hierarchical team and individual memories to retrieve only currently valid information. The system reduces outdated retrieval and improves downstream question-answering quality in collaborative LLM agents. Builders care because the framework prevents agents from answering with outdated or conflicting memories that no longer align with current team consensus.

Summary by a local model.

Source: arxiv-cl · 2026-09-28

There is more to code review than (automatable) detection

The article argues that coding agents have crossed a capability threshold where traditional human code review is no longer necessary. The author counters this by identifying four overlooked aspects of peer review, including the signal of human confusion and the ability to notice missing elements. Builders care because the article claims agents can serve every stated goal of code review at lower cost and higher throughput.

Summary by a local model.

Source: hn · 2026-09-26