Introducing Rubrics: Build Agents that Evaluate and Correct Their Work
Deep Agents' new RubricMiddleware adds a self-eval loop: set a rubric, configure a grader, and the agent corrects its own output before returning it.
AI Daily Recap
19 articles · 5 categories
The finishable daily brief
What happened in AI — Aug 25, 2026
Tuesday, Aug 25, 2026
19 articles · 5 categories
read top to bottom · then stop
In 30 seconds
- LangChain shipped RubricMiddleware for Deep Agents self-evaluation, plus a LangSmith Engine update that detects agent issues 2x better.
- AWS added MCP Apps to OpenSearch Service so agents can return interactive traces alongside their text responses.
- QiAnXin disclosed a critical remote-code-execution flaw in DeepSeek's agent harness.
- OpenAI's CFO laid out the compute economics behind scaling intelligence, while its new Jalapeño inference effort posted early speed/efficiency results.
- Low-cost, open-weight Chinese models led by DeepSeek keep gaining share on US model-hosting platforms.
Agent tooling converged on one theme today: catching an agent's own mistakes before they reach the user. LangChain shipped a self-eval rubric middleware and a LangSmith Engine update that detects agent issues twice as well, AWS added MCP-based trace observability to OpenSearch, and an independent Show HN tool renders agent traces like a navigable JS bundle.
Security cut the other way. QiAnXin disclosed a critical remote-code-execution flaw in DeepSeek's agent harness, and OpenAI banned a Russian influence operation running on its models — a reminder that the same agent infrastructure getting easier to observe is also an expanding attack surface.
Agent Evals & Observability 4 items
Agent tooling is converging on self-correction: LangChain shipped a rubric-based self-eval loop and a 2x-better issue detector, while AWS and an independent tool add trace-level observability to agent runs.
LangSmith Engine Improves Agent Issue Detection by 2x
LangSmith Engine now detects agent issues over 2x better and proposes stronger fixes, with Slack/Linear workflows and self-hosted deployment support.
Agentic observability with Amazon OpenSearch Service MCP Apps
OpenSearch Service's new MCP Apps let a single local MCP server return interactive traces alongside an agent's text response, moving debugging from alert to root cause in one hop.
Show HN: Unbox-AI – Visualize your AI traces like a JavaScript bundle
An open-source tool renders agent trace JSON as a navigable bundle-style visualization instead of raw text a builder has to paste back into another agent to parse.
Agent Memory, Environments & Tooling 4 items
Builders are formalizing the substrate agents run on — skill libraries, self-correcting knowledge stores, and synthetic task environments — plus a push to give every agent its own sandboxed compute.
Using skills with Deep Agents
Deep Agents CLI now discovers, loads, and executes reusable skills dynamically, cutting redundant context per run.
Building Self-Correcting Memory in OpenWiki
OpenWiki tags claims with evidence so it can detect when its own knowledge base has gone stale and self-correct, instead of silently hallucinating from outdated context.
How We Build Agent Environments & Tasks
LangChain's synthetic-environment pipeline splits into a spec-generation step, a spec-to-task step, and a shared world spec — a template for building agent eval environments at scale.
We gave every agent a computer
A new project gives each agent its own sandboxed compute environment rather than sharing a host shell, aimed at safer parallel agent execution.
Coding Agents & Developer Tools 3 items
Coding-agent tooling keeps fragmenting into specialized pieces: a model-agnostic terminal agent, local context retrieval for code agents, and admin controls for enterprise ChatGPT/Codex deployments.
Show HN: TheGitAI – A model-agnostic coding agent for your terminal
A terminal coding agent that swaps between models rather than locking into one provider's CLI.
Cortex – Local context retrieval for AI coding agents
Cortex retrieves relevant local codebase context for coding agents without shipping the whole repo into the prompt.
Introducing the Admin plugin for ChatGPT Work and Codex
The Admin plugin lets IT teams analyze workspace usage, manage members and permissions, and adjust limits for ChatGPT Work and Codex from one console.
Compute Economics, Models & Enterprise Adoption 4 items
OpenAI's CFO and a new inference effort both made the case that compute economics — not just model quality — now drive who wins, as low-cost Chinese open-weight models keep gaining share on US platforms.
The full stack behind abundant intelligence
OpenAI's CFO argues that compounding gains across chips, compute, models, and products — not any single breakthrough — are what's driving intelligence cost down and scale up.
Jalapeño's first results show industry-leading speed and efficiency in AI inference
OpenAI's early results for a new inference effort called Jalapeño claim industry-leading speed and efficiency, positioning it against other inference providers.
DeepSeek leads surge in low-cost Chinese open-weight models on US platform
Low-cost, open-weight Chinese models led by DeepSeek are gaining traction on US model-hosting platforms, per SCMP, pressuring the pricing floor for proprietary frontier models.
Bain & Company joins the Claude Partner Network as a Global Premier partner
Bain becomes a Global Premier partner in the Claude Partner Network, building on its rollout of Claude to 19,000 employees — enterprise consulting standardizing on a single model vendor.
Security, Privacy & Safety Policy 4 items
Today's security news cuts across the stack: a critical RCE in DeepSeek's agent harness, a confidential-computing pitch for private cloud inference, and platform-level responses to both an influence operation and AI's wellbeing impact.
QiAnXin Discloses Critical Remote-Code-Execution Flaw in DeepSeek Harness
Security firm QiAnXin disclosed a critical RCE vulnerability in DeepSeek's agent harness, the coding/tool-execution layer many DeepSeek-based agents run on top of.
New technology lets cloud AI models process your data without privacy leaks
PlugClaw pitches confidential computing for cloud AI inference so a provider can process your data without being able to read it.
Disrupting a new covert influence campaign from Russia
OpenAI banned Russia-origin accounts that used its models to run a fake Israel-based think tank and a Western-critical “sovereignty” index.
Funding better evaluations of AI’s impact on wellbeing
Anthropic is funding $5M in independent research grants to build better evaluations of how AI use affects users' wellbeing.
You are caught up for this edition