SATURDAY · AUGUST 22, 2026 · ISSUE 4 · YESTERDAY
Daily News — 2026-08-22
0 papers indexed on arXiv·~6,600 packages released on PyPI·26 papers surfaced by Hugging Face
None of it is in your agent's weights.
Spotlight
Wrapping static benchmarks instead of replacing them is a pragmatic bet, and the held-out performance gains suggest it pays off.
LLM agents are only as good as the environments they train in, and most training environments are built once and never updated. EnvHarness attacks this problem from an unusual angle: rather than generating new…
A one-sentence stub: `dspy-ai` is now just an alias — replace it with `dspy` in every dependency file you own.
The dspy-ai package on PyPI is now a stub — nothing more than a forwarding address. Its entire published description is a single sentence telling you to install dspy instead.
Papers
Wrapping static benchmarks instead of replacing them is a pragmatic bet, and the held-out performance gains suggest it pays off.
EnvHarness uses programmable plugins to dynamically reshape static environments and target agent weaknesses in reinforcement learning.
Below 50% for the best agent tested — and the failure taxonomy explains why more context can make scientific code repair worse, not better.
SWE-bench Science tests coding agents on scientific software repair, revealing failure mechanisms and mixed effects from scientific guidance.
Multi-turn simulation as a feedback generator, not an evaluation endpoint, is the specific mechanism that keeps skill evolution gradients from decaying after the first round.
SkillEvo uses multi-turn interaction feedback and active governance to prevent skill evolution gradients from stalling in LLM agents.
Explicit modularity metrics—not more LLM calls—are what make iterative architectural refinement converge rather than spiral.
Repo0 uses a dual-graph architecture and modularity-guided evolution to generate complete software repositories from natural-language requirements.
Seed variance swamps every accuracy effect; the real gains — reasoning language, budget discipline, format compliance — live where benchmarks don't look.
SFT shifts reasoning into a low-resource language without accuracy loss, but RL is needed to fix the formatting and leakage defects SFT leaves behind.
Eximo's key result: distilling VLM-guided trajectories back into the base VLA beats keeping the VLM live at evaluation time, and beats RL-only finetuning given the same data budget.
EXIMO fine-tunes large vision-language-action robot policies via VLM-guided exploration, imitation on orchestrated data, and residual off-policy RL.
10 more tool picks in this edition
Every pick in the Wire gets the same treatment: read, verified, and given a written verdict.
A three-stage decomposition that separates document exposure, QA alignment, and capability recovery—and earns its gains by measuring each stage separately.
IAR post-trains LLMs in three stages—inject, align, recover—to internalize document knowledge for retrieval-free QA without sacrificing general capability.
A frozen LLM rewriting its own scaffold and its rewriting strategy yields real gains on moderate tasks, with the failure modes named precisely rather than hidden.
Hierarchical self-modification evolves execution harnesses around frozen LLMs, yielding strong gains on moderate tasks but bounded by feedback quality.
A benchmark that finally separates 'long video' from 'context-rich video' — and proves current models cannot reliably handle the latter.
A Japanese long-form video benchmark for evaluating narrative evolution and cultural reasoning, built with native-speaker verification.
Tools & packages
A caching layer for browser-agent knowledge that compiles site exploration into reusable CLI adapters - the concept is sharp, but the README skips the hard questions about staleness and internals.
Browser infrastructure that learns from interactions to reduce token spend by up to 90%.
A one-sentence stub: `dspy-ai` is now just an alias — replace it with `dspy` in every dependency file you own.
A framework for programming language models by composing declarative modules rather than writing prompt strings.
A browser built for agent-human cohabitation, with a code-over-CLI execution model that cuts tool-call overhead — macOS only for now.
A browser built for sharing logged-in browser state with AI agents like Codex or Claude Code, with zero config required.
Parallel git worktrees per agent, with diff annotation and a mobile monitor, is a real workflow improvement over juggling terminal tabs.
Orca is a desktop and mobile ADE for running a fleet of parallel coding agents with your own subscription.
Ten formatting rules that force a coding assistant to lead with the action and drop the filler — the whole value is in the discipline, not the technology.
Coding-agent skill that surfaces answers up front instead of burying them — built around ADHD-friendly output conventions.
A pre-execution policy layer for twelve agent harnesses that blocks dangerous tool calls locally, no account required — enforcement on custom runtimes still needs your own hook.
A runtime layer for coding agents that catches loops, dangerous actions, and secret leaks before they become incidents.
9 more paper picks in this edition
Every pick in the Wire gets the same treatment: read, verified, and given a written verdict.
A Rust MCP server that compiles agent session observations into a git-versioned markdown wiki and hands context forward across agent switches, with no vector database required.
A repository providing long-term memory for agent coding CLIs and handoff between different agent vendors.
A continuous reward framework for agent trajectories that turns logprob distributions into actionable selection and progress signals.
A general-purpose framework that gives fine-grained feedback to agents across coding, robotics, and medical benchmarks without additional training.
A local memory bus for multi-agent workflows — the idea is sound, but the retrieval quality that makes or breaks it isn't shown.
A local memory hub that gives AI agents shared, persistent context so every agent retains the same information about you.
A hardware-first model selector that makes its speed estimates auditable and lets users replace them with real measurements via a built-in benchmark loop.
One command checks hundreds of models and providers to surface which ones your hardware can actually run.