惯性聚合 高效追踪和阅读你感兴趣的博客、新闻、科技资讯
阅读原文 在惯性聚合中打开

推荐订阅源

D
DataBreaches.Net
GbyAI
GbyAI
aimingoo的专栏
aimingoo的专栏
freeCodeCamp Programming Tutorials: Python, JavaScript, Git & More
月光博客
月光博客
大猫的无限游戏
大猫的无限游戏
M
MIT News - Artificial intelligence
腾讯CDC
博客园 - Franky
Engineering at Meta
Engineering at Meta
C
Check Point Blog
T
The Blog of Author Tim Ferriss
有赞技术团队
有赞技术团队
Microsoft Azure Blog
Microsoft Azure Blog
MyScale Blog
MyScale Blog
I
InfoQ
Blog — PlanetScale
Blog — PlanetScale
P
Proofpoint News Feed
The GitHub Blog
The GitHub Blog
N
Netflix TechBlog - Medium
Last Week in AI
Last Week in AI
S
SegmentFault 最新的问题
Cyber Security Advisories - MS-ISAC
Cyber Security Advisories - MS-ISAC
WordPress大学
WordPress大学

TestingCatalog

SpaceXAI gearing up for upcoming Grok 4.5 release ByteDance set to launch Seedance 2.5 with 3-minute output Meta prepares Scheduled tasks for Meta AI users on web Google tests new Gemini Inbox section for Workspace triage OpenAI might be preparing GPT-5.6 for next week's release Mistral releases Leanstral 1.5 model for proof engineering xAI debuts Grok Voice Agent Builder for Enterprises Condense launches proxy to cut AI coding agent bills by 66% Vellum adds agent-to-agent AI collaboration for Slack Early look at Anthropic's Claude Science app for researchers Google launches Nano Banana 2 Lite and Gemini Omni Flash Google might be testing Gemini Flash upgrade on LM Arena Anthropic may impose KYC restrictions for Fable 5 access Anthropic launches Claude Sonnet 5 model on Claude and APIs NoimosAI launches Creative Agent for brand assets Apify lets AI Agents pay via Coinbase x402 for web tools Bloome launches chat platform for AI agent teams Meituan launches LongCat-2.0 1.6T parameter model Cursor releases its iOS app for vibe coding on the go OpenAI prepares upgraded Office controls for Codex OpenAI tests gifting Codex credits as new growth strategy Microsoft launches MAI-Code-1-Flash on GitHub Copilot Google adds Computer Use to Gemini 3.5 Flash Google tests notebook collections for NotebookLM OpenAI launches GPT-5.6 Sol preview for select partners Microsoft adds Copilot finance tools to Excel for M365 users DeepReinforce releases Ornith-1.0 open-source coding models Gemini to get voice dictation and Magic Pointer on desktop Meta launches AI glasses with three new styles from $299 Anthropic launches Claude Tag on Team and Enterprise plans
OpenSquilla lets AI agents organize their own skills
Nero Soares · 2026-06-05 · via TestingCatalog

Most agent projects are still racing to field a smarter chat loop or a longer list of tools. OpenSquilla, an Apache-licensed, self-hostable runtime, is making a quieter bet: that the next round of efficiency comes from the harness rather than the model. Its founding idea was cost-aware routing, scoring each turn and sending trivial work to cheap models while reserving heavier reasoning for tasks that warrant it. That is becoming table stakes. The part worth watching now is MetaSkill, the project's attempt to let an agent organize its own capabilities rather than rely on hand-written workflows.

OpenSquilla

The premise is the combinatorial problem facing every maturing agent: writing a single-task skill is trivial, but composing hundreds of community skills into something reliable collapses into guesswork once real complexity arrives. MetaSkill answers with a meta-protocol, a markdown spec that tells the model how to discover, rank, and compose atomic skills, declaring the resulting workflow in a structured header that the runtime validates before anything runs. A goal described in plain language becomes an inspectable, replayable execution chain rather than a one-off prompt.

Tonight, as promised 🦞

That volcano plan this morning? Not a chatbot — it was
MetaSkill, OpenSquilla's self-organizing skill protocol.

You describe the goal in plain words. It discovers, picks,
and composes the right skills into a real, safe workflow —
and it can even write… pic.twitter.com/f1ksWydMp5

— OpenSquilla (@OpenSquilla) June 1, 2026

The runtime ships with ready-made workflows for jobs such as research-to-report and project planning, and during idle time, it revisits its execution traces, distills recurring patterns, and drafts candidate workflows. The catalog grows in the background. That is also where the open questions sit. The headline savings figures are the project's own benchmarks, not independent results, and machine-composed workflows raise obvious reliability and safety concerns that the design seeks to contain through proposal gates, tool allowlists, and syscall-level sandboxing.

Strategically, OpenSquilla is positioning against heavier agents such as OpenClaw, even shipping migration tooling, while aligning with the wider move toward portable skill specifications. For teams running long-horizon agents where token bills compound, the pitch lands squarely at the orchestration layer. Routing, tiered memory, and the first MetaSkill capabilities sit in the public releases today; a further iteration looks set to follow from active development. If orchestration rather than model size proves the durable lever, projects like this reframe where the real moat lies.