Rubric Devs

Updates from the developers building at @RubricLabs

[Ted Spare @TedSpare]

Agent harnesses take strong opinions about many things:
- tool-calling
- memory
- subagents

A critical one is compaction.

We recently saw that by using a native compaction strategy, 5.6 Sol did 3x better on ARC-AGI-3 at 1/6 the tokens → ~18x token efficiency from harness diff.

Curious whether the same effect applies to other domains, we ran a small experiment and found Pi passed some TerminalBench-2.1 tasks with 30% fewer tokens when using native compaction.

There is so much low-hanging fruit to harness engineering, and since it's just code, it's accessible to a huge range of teams.

[dexter @dexterstorey]

Memory is still an unsolved problem.

SKILL.md is for procedural memory.

MEMORY.md is for everything else.

But we don't have good patterns for how and when to use them, so agents store arbitrary and partial information.

And we end up with:
- lossy history
- stale facts
- bad instincts

"the user prefers pushing to main over PRs"
(true in context, not globally)

"the user is building a todo list app with firebase"
(stale)

"use external dependencies where possible"
(written by a sub agent, treated as a law)

[Sarim Malik @sarimrmalik]

Really enjoyed the animation work we did for this post, converging on a nice design language.

[Rubric Labs @RubricLabs]

What is an agent harness?

https://t.co/TN4TNKhqkf

[Laurence Liang @LaurenceLiang1]

this past June, we hosted an AI safety hackathon at Omega Labs by @HorizonOmega_ , Montreal's newest AI safety space.

supported by @redwood_ai , @RubricLabs , @rootlyhq and with Brev compute by @nvidia , we received 80+ signups for a week of AI safety research. highlights in🧵!

[dexter @dexterstorey]

when someone gives us feedback, we put up a PR which automatically publishes a release-candidate on NPM and pings the person to test it.

it's kinda like frontend preview urls - but for a CLI.

thanks to @pedroapfilho, @lennardeth and @TedSpare for the 🔥 feedback

[Rubric Labs @RubricLabs]

E2B has deployed 1 billion AI sandboxes and worked with 94% of Fortune 100 companies.

We had the chance to work with them and learn about their process. We wrote about it in a blogpost.

https://t.co/OokYa8wWrY

[Rubric Devs @RubricDevs]

how many agents are you running?

[dexter @dexterstorey]

Introducing TOKENMAXX

We built a CLI that merges your usage across all your Claude Code or Codex subscriptions.

We’ve been doing lots of overnight runs that burn through multiple accounts. Now we don’t need to babysit them.

Try it → bun i -g tokenmaxx

[Rubric Labs @RubricLabs]

Challenging the Chatbot

We’ve been exploring UX patterns for AI outside of the chat interface.

Chat has become the de facto standard for AI. What if we instead embed intelligence into the interfaces we already know?

https://t.co/Auk0mbxgMa

[Laurence Liang @LaurenceLiang1]

@nvidia is joining us as a sponsor along with @redwood_ai , @RubricLabs and @rootlyhq !

a special thanks to all our sponsors for the week-long AI safety (control) hackathon, launching this Thursday both irl in Montreal and online

signups still open, details in thread!

[Rubric Devs @RubricDevs]

Something new this way comes.

Can you guess what it is?

[Rubric Devs @RubricDevs]

Agents now routinely write working code.

The software still feels subtly wrong.

We lay out how we define “done” for software in the agentic era.

[Rubric Labs @RubricLabs]

New Post: Contract Engineering

AI agents ship broken software. Tests pass, the app deploys, and it’s still wrong.

We’ve been writing contracts: Requiring proof that every event fires in order for every use case.

Once all contracts are filled it ships. https://t.co/zJxpcoWnLo

[Rubric Devs @RubricDevs]

New icon dropped 👀

https://t.co/Qwa1xMyeuR

[Rubric Labs @RubricLabs]

New post: Unblocking Agents

Agents are capable of building great software. What stops them: permissions, credentials, infrastructure, and identity.

We stood up a parallel org called Rubrot and gave agents admin permissions, hardware and credit cards. https://t.co/37NIbmn15y

[Rubric Labs @RubricLabs]

New post: Primitives over Pipelines

The AI systems most teams are shipping today were designed for dumber models. Now that frontier intelligence can follow instructions, reason, and self-correct the move is to give agents primitives and let them cook.

https://t.co/8j0rdSnHIz

[Valth @valthvisuals]

SOUND ON 🔊 Sound design breakdown

[Ted Spare @TedSpare]

GPT-5.2 from @OpenAI launched yesterday and we wanted to test it.

We built a playground to test its coding + UI capabilities and set a $1000 API limit.

Go nuts: https://t.co/AGRJ5EK336

[Rubric Labs @RubricLabs]

AGI is on hold.

Long live Structured Outputs.

[Rubric Devs @RubricDevs]

we did it

new high score

[Erik Langille @eriklangille]

Introducing techtris - Play tetris with your favorite AI startup company logos

featuring @Replit, @RubricLabs, @Figure_robot, and @tryprofound (the final techtris piece boss)

[Rubric Devs @RubricDevs]

@dparksdev @RubricLabs ok

BE → https://t.co/OXB3stFfjK

FE → https://t.co/Xq8hcrpevz