While building my own agentic/coding harness, I noticed how differently agents structure the fundamental message block. Users provide text or images as input (in coding agents, audio is usually transcribed to text first). Tool calls come from the assistant, but providers differ on where tool results go:
OpenAI uses a dedicated `role: "tool"`, while Anthropic nests them in user messages.
If you’re building a harness, start by defining your primitives (UserMessage, AssistantMessage, ToolMessage), what inputs each accepts, and a strict schema for each.
Here’s the primitive structure from my harness for reference.
Too many factually inaccurate points are made here. One should get his facts right!
1. "No one is going to really understand it". Data says otherwise: 3.1 million reported speakers, 18 universities worldwide, 100+ academic journals.
2. "Sanskrit was always the language of very few in India since ancient times" - again, the data says otherwise.
3. "could never develop india scientifically" - https://t.co/PXsWSmERnt - Indian Math book in Sanskrit discussing calculus.
4. "there is nothing technical about Sanskrit" - We have shown near-Turing completeness with Sanskrit Sandhi rules.
Just tried https://t.co/U2iMlafGyO. Absolutely loving it, I'm using it like cursor, with inline edits and threads connected to claude account.
Didn't have to setup, it instantly detected the default claude account in device.
Great job @arpit_bhayani
100 AI Styles in 100 Days
Day 77/100:
📷 Art style: Indian Mythic Folk Art
(Indian mythological art + folk illustration + flat geometric shapes + vintage editorial/storybook)
📷They set his tail on fire. He set the whole city on fire. 🔥🐒
Inspired by the ancient Indian epic, Ramayana.
📷 Created with Seedance2.5 on @ImagineArt_X@imagineart_creo
@arpit_bhayani How Cerebras Built Its Enterprise Knowledge Base https://t.co/TUaEaXl7zT
Combining knowledge base and in-repo knowledge might be the clear trajectory towards scale I think.
But i also see RAG going away in favor of just letting agent grep over data with clean metadata.
@X@nikitabier
Getting elaborate phishing attempts trying to gain access like this, with another account claiming they reported me for wrong activity.
Then claiming to contact a fake @moderator_yale account for resolution.
Keeping track of agent-led researches and experiments have become harder with time, especially with autoresearch loops and agent-led investigations.
I have built autoexp — A local-first experimentation workspace for AI Agents.
Try it out at https://t.co/CNmZitYgki