Tarushi

APM @AmericanExpress🌷
Building Recap 🛠️
Being in motion 🤸🏻‍♀️

[Akshay 🚀 @akshay_pachaar]

Don't train the model, evolve the harness.

I read a brilliant blog post from Hugging Face where they took a frozen open model scoring 0% on a hard legal agent benchmark, left its weights alone, and let an automated loop rewrite only the code around it.

That code layer is the harness, the runtime wrapper that feeds the model context, runs its tool calls, and decides when a run ends.

By the time the loop finished, the system had essentially matched Sonnet 4.6 on the benchmark's headline metric, at roughly 7x lower cost per task. Zero weights changed.

The gain existed because of where the model was failing. The judge only grades files saved in the right place under the exact requested filename, and the model kept doing the legal analysis correctly, then saving it under the wrong name, dropping it in a scratch folder, or never writing it at all.

So the 0% was never measuring legal reasoning. It was measuring the harness.

Hand-tuning that layer is slow and model-specific, so they automated it. A Claude proposer adds exactly one mechanism per iteration, and an outer loop keeps it only if it clearly beats the current best, so accepted mechanisms compound.

What the loop discovered says a lot about where agents actually fail.

�� The biggest single gain was file handling, not intelligence. An automatic step that lands the deliverable exactly where the judge expects it beat every prompt change, with zero extra model tokens.

→ Code fixes transferred across models, prompt playbooks did not. The same harness lifted a smaller model from the same family by 14 points, but the tuned prompts hurt a different model family on tasks it could already finish.

→ The harness mattered more than anything else. Same model, same judge, same tasks, and five different harnesses scored anywhere between 3.5% and 80.1%.

The gains do eventually flatten, and the remaining misses look like real capability gaps. At some point the wrapper runs out of tricks and the model has to carry the work.

But the lesson holds. A benchmark score measures the model and its harness together, and until the harness is fixed, it's impossible to know which one failed.

I highly recommend reading this: https://t.co/3ZIeKhsngn

I also wrote a deep dive on agent harness engineering a while back, covering the orchestration loop, tools, memory, context management, and everything that turns a stateless LLM into a capable agent.

The article is quoted below.

[Tarushi @floatinAnanas]

@zeelprajapati__ @GemsOfCricket Not me saving my face cause he came to my college's graduation ceremony.

[Tarushi @floatinAnanas]

@Prashant_1722 Can I get the code

[Tarushi @floatinAnanas]

@YoJacq SORA

[Tarushi @floatinAnanas]

@aryanXmahajan VAULT

[Aryan Mahajan @aryanXmahajan]

i created a fully open source automation arsenal of 3,258+ AI workflows...

(the exact arsenal agencies charge $100K+ to develop)

here's what you get:
- 3,000+ n8n workflows including AI ad creative agents
- Intelligent Context-Driven Claude MCP Sales Agents
- Claude MCP agents for complete LinkedIn automation
- AI UGC system that saves $247 per video (5-minute creation)
- 100+ AI prompts that generate world-class copy
- AI Viral Content OS (complete content empire)
- Email/SMS automation sequences that convert at 18%+
- LinkedIn Empire Blueprint (exact system that built 30K+ followers)
- 7,000+ Make workflow automation templates
- Enterprise AI prompts (100+ commands that outperform $500/hr developers)

basically everything you need to automate your entire business while your competitors still hire expensive agencies

this complete AI Business Vault puts you 2+ years ahead of the competition

Like + comment “VAULT” + repost (must be following so I can DM you)

[Tarushi @floatinAnanas]

@its_rizzwan @anushkag319 @AmericanExpress Haha, hey @anushkag319 ✌🏻

[Beerbal @Shikshita_Juyal]

Seeking a Rockstar Growth Intern 🚀

I’m on the lookout for a sharp, passionate Growth Intern to join me as we scale. This is a part-time, remote role starting at ₹15K/month (with potential to scale to ₹30–35K based on results over time).

No experience? No problem. We’re not-

[Tarushi @floatinAnanas]

@ishanjain_ Peers jab kutte honge toh yehi hoga na😂

[Tarushi @floatinAnanas]

@radika_agarwal Love this, AI for curiosity is always intriguing to think about - the right nudges at the right time really helps a person to connect diff topics they've learnt about

[Tarushi @floatinAnanas]

I can't stop laughing, someone vibecoded the Coldplay Concert xD💃🏻🪩
https://t.co/2lMPcrMSdX

[Tarushi @floatinAnanas]

1mo into doing PM stuff - I already want to go back to coding 1 day and chilling for the entire week 🤡

[Tarushi @floatinAnanas]

@Amani_Chowdhry 👋🏼

[Tarushi @floatinAnanas]

@aaditsh AI

[Tarushi @floatinAnanas]

My feed is showing me some "n8n video" and dance on "shake it to the max" back to back😭
Dude, social media genuinely needs modes - tech mode, dance mode, literature etc. I like them all, but not all at once

[Shubham Saboo @Saboo_Shubham_]

I built a Multi-Agent Product Launch Intelligence app using Firecrawl and Agno.

It has 3 AI Agents working together:

- Competitor Analysis Agent
- Market Sentiment Agent
- Launch Metrics Agent

100% Opensource Code with step-by-step tutorial.

[Aditya Thakur @createdbytango]

Hosted my first talk today @tap_invest on curiosity and taste!

[Tarushi @floatinAnanas]

@aaditsh AI

[Tarushi @floatinAnanas]

@UtsavChindalia This looks so cool🔥🔥

[Tarushi @floatinAnanas]

Ek part time boyfriend toh mai bhi deserve karti hu, just to soft launch him through Ghibli style picture xD🍃