The Archive

James Shore argues AI coding agents must reduce maintenance costs inversely to productivity gains or risk long-term debt; doubling output without halving maintenance costs creates net negative ROI.

SLIM: dynamic skill lifecycle management for LLM agents enabling non-monotonic skill activation based on task and stage.

Shepherd: runtime substrate for meta-agents with formalized execution traces in Lean, enabling 5× faster forking and state replay.

WildClawBench: native-runtime benchmark of 60 real-world, long-horizon CLI agent tasks (8+ min each) for LLM/vision-language agents.

AI agents need formal SE practices (testing, staging, adversarial eval) beyond on-the-fly synthesis for high-stakes deployment.

AssayBench: benchmark for LLMs and agents on virtual cell phenotypic screening combining textual inputs with diverse cellular outputs.

Pentesting agents evaluated on real-world targets show current benchmarks miss complexity and strategic decision-making required in practice.

Framework enables visual-native multimodal search agents with on-policy data evolution and persistent visual evidence reuse.

MaD Physics benchmarks agents on resource-constrained scientific discovery with real measurement trade-offs and planning.

Anthropic's Claude Platform reaches general availability on AWS with managed agents, code execution, web search, batch processing, and same-day feature parity with native API.

I checked out GitHub after Anthropic's announcement last week and came across their new financial services reference repo (github.com/anthropics/financial-services). It packages 10 pre-built workflow agents for financial services firms you can run them through the Claude Cowork plugin or via the Managed Agents API, which is a useful bit of flexibility depending on how your stack is set up. The 10 agents cover a decent spread of the typical pain points: * **Pitch Agent** \- builds fully branded pitch decks from comps, precedent transactions, and LBO analysis * **Meeting Prep Agent** \- draft...

This was a really good talk, especially for anyone who's built things like the Karpathy wiki, Serena, or SQLite databases as memory for Claude. For any senior devs out there, are you spotting the solutions already implemented in distributed systems being reused? If many agents are working in parallel, how do you get them from stepping on each others toes? I can imagine logical clocks, consensus, deduplication, idempotency, and eventual vs causal consistency being applied. If you're on the Anthropic team, I'm curious how much different distributed systems algos were experimented with.

Reddit discussion questioning OpenAI's hiring of OpenClaw creator and value of agents initiative; post contains factual error later corrected.

Claude Max still produces hallucinations causing production failures; autonomous agents unsuitable for unsupervised deployment without guardrails.

Agentic framework embeds LLM agents in SCIP solver harness to auto-generate and benchmark constraint handlers.

Selective imitation learning framework enables agents to abstain from acting when demonstrations are uninformative under dynamics shift.

CIVeX verifies causal effects of tool-use actions in LLM agents via structural causal queries and identifiability checks.

MCP-Cosmos integrates world models into Model Context Protocol agents to bridge planning-execution gap via predictive task automation.

Framework maps LLM reliability techniques (retry, voting, self-consistency) to Shannon coding theory operators as stochastic channel reliability methods.

First comprehensive survey unifying token economics for LLM agents, framing tokens as production factors and analyzing computational-economic trade-offs.

Reddit discussion questioning the long-term value of Anthropic's Claude Certified Architect credential as AI agents automate architecture decisions.

Like many others here, I got frustrated with managing all my different claude/codex sessions, so i built Pokegents, which is an open source multi-agent workspace for coding agents. It has a Pokemon-themed dashboard/chat interface plus a local orchestration server for managing agent sessions (currently supports Claude Code in iTerm2, plus Claude and Codex through ACP-based chat runtimes), persistent agent identities, mcp messaging between agents, notifications, session cloning, and more. This was mostly a vibe-coded side project, but I've been using it constantly in my day-to-day workflow as ...

Study shows expanded context windows degrade multi-agent cooperation in LLMs across 7 models; mechanism is eroding forward-looking intent rather than increased distrust.

Bash is one of the most flexible and powerful interfaces exposed to AI agents. In the right system, a model that emits grep, curl, tar, or a shell pipeline is... Bash is one of the most flexible and powerful interfaces exposed to AI agents. In the right system, a model that emits , , , or a shell pipeline is producing an executable action that can read files, mutate a workspace, open network connections, and chain tools together. For the NVIDIA AI Red Team, this makes command generation a useful research target. If smaller language models can be guided… Source

Proposes action-credit RL for CLI agents using structured command attributes and selective observation over filesystems.

Susceptibilities technique extends neural network interpretability to deep RL agents, revealing parameter-space development undetectable from policy analysis alone.

Measures optimal timing for clarification requests in long-horizon agent execution via injection framework across 4 dimensions.

DRIP-R benchmark evaluates LLM agents on real-world retail policy ambiguities with multiple valid interpretations, addressing evaluation gaps in agent robustness.

Autonomous agent oversight reveals endogeneity: non-affine approval functions needed to screen dishonest agents violate truthful reporting conditions.

PhoneSafety benchmark (700 examples) distinguishes genuine safety understanding from task failure in phone-use agents via fine-grained outcome categorization.