@palanthos

AI agents, explained. Research and practical guides on tools, workflows, and building with agents. Building the AI Agent Era.

palanthos.com
Joined April 2026
  • Tweets8
  • Following22
  • Followers3
A convincing final screen can hide an instruction mismatch. AgentHorizon (Oct 8 preprint) tests judges on paired human recordings and near-misses. Our takeaway: match actions to the request, not just the output. arxiv.org/abs/2610.11050

AgentHorizon: Evaluating Agentic Judges for Long-Horizon Computer-Use Tasks

Computer-use agents are capable of completing complex tasks, increasing the use of automatic judges to determine success, either for training or for evaluation without human involvement. Despite...

arxiv.org
Copilot review APIs, announced Oct 2: request != completion or approval. GitHub says Default -> Balanced took effect Sep 28, 2026; explicit Lite was respected. Check effort on the submitted review. github.blog/changelog/2026-1…

Copilot code review: API support and new default effort level - GitHub Changelog

You can now request a GitHub Copilot code review through the REST and GraphQL APIs and set the review effort level for each request. Balanced is also now the default…

github.blog
2
A voice model's low word error rate measures intelligibility, not naturalness. Open TTS separates these signals. For a bilingual agent, English scores cannot settle the choice: compare target-language errors and matched listening tests. huggingface.co/blog/open-tts…

Open TTS Leaderboard: Scalable Evaluation for Multilingual Text-to-Speech and Voice Cloning

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

huggingface.co
1
METR's Feb 2026 coding study update: estimated time savings came with intervals spanning no effect. Developers also withheld tasks they wanted AI for. Randomizing submitted tasks cannot fix missing ones. Measure selection, not just speed. metr.org/blog/2026-02-24-upl…

We are Changing our Developer Productivity Experiment Design

Our second developer productivity study faces selection effects from wider AI adoption, prompting us to redesign our approach.

metr.org
2
We've partnered with @OpenAI to bring the new Sign in with ChatGPT experience to Nous Portal. Log into Nous Portal with ChatGPT to use your plan in Hermes Agent, with full visibility and controls in ChatGPT settings.
162
219
65
3,070
285,153
GPT-6.1 Sol: $2 input / $10 output. Astra: $10 / $50. USD per 1M tokens; standard, short context, checked Oct 1. Cheaper rates do not establish cost per accepted task. Retries and review still count. developers.openai.com/api/do…

Pricing | OpenAI API

Pricing information for the OpenAI platform.

developers.openai.com
6
OpenAI's dots docs draw a useful boundary: proactive research can read connected information, but cannot send messages or change apps. Follow-up actions still need applicable permissions. Documentation, not our product test. learn.chatgpt.com/docs/dots/…

Tasks and memory | ChatGPT Learn

Understand assigned work, recurring tasks, and your dot's context

learn.chatgpt.com
4
A claim can be true and still cite the wrong source. ProvenanceGuard studies this failure in multi-tool agents. Check factual support, attribution and useful recovery separately. Medical-trace research; not independently replicated here. arxiv.org/abs/2606.18037v3

ProvenanceGuard: Source-Aware Factuality Verification for...

Tool-using LLM agents increasingly use the Model Context Protocol (MCP) to answer from heterogeneous evidence sources, including search, APIs, databases, clinical records, and formulary tools....

arxiv.org
7
Load more