Averaging your LLM judges gives the weakest one the same vote as the best. On a 200-item gold set, you can’t tell them apart.
AGORA (arXiv, Nov 3) uses item-response theory to learn each judge’s reliability without labels.
How to try it: arxiv.org/abs/2611.05556
2
679
Charging platform fee by e-comm by platform is like if a local shopkeeper ask customers a portion of rent for the shop
181
Coding agents are very good at building what you ask for.
The gap between what you specify and what you expect is where the agent starts making assumptions.
More assumptions → more rework → more iterations → more tokens.
The clearer the specification, the closer the output.
🤖 Made with AI
1
165
Replying to @nikks_techie
It should always be tools/backend driven where agent only controls and updates the agent state
1
52
Replying to @nikks_techie
I usually have state checked in the tool itself. For instance, when one tool is finished and it alters some state, and the agent decides which tool to use next based on the workflow, the next tool itself will check the state.
17
Replying to @nikks_techie
Clearly define:
- what the purpose of each tool is
- what the inputs and output map are
- when failure is considered
- when success looks like
is usually works really well in practice
1
1
102
Replying to @Abhishekcur
Fair points. I think the answer is to use AI more, experiment, notice what consistently works, and build processes around those patterns—just as we’ve always done with new technology.
1
47
Many people want AI to handle their thinking, execution, and everything in between—which surrenders their leverage. I think, AI should execute and retrieve information for me, but I keep the thinking.
1
183
An LLM is a next-token generation machine whose response is driven by system prompt + context + current query. If the requirements aren’t clear enough, you’ll just keep burning tokens without getting good results.
Many times users aren’t clear on their requirements for an LLM. They expect surprisingly good results even when they themselves don’t know what output they’re looking for.
1
2
283
Is it just me, or does Codex hit its usage limits much faster than Claude for the exact same task under identical conditions?
1
222
Two agents can use the exact same LLM and still be radically different systems.
Give one:
full filesystem + network + shell
Give another:
sandbox + scoped files + restricted tools
Same intelligence. Different capability. Different blast radius.
Agent = Model + Harness
2
2
469
The interesting part of “agentic video” isn’t that an AI can watch a 90-minute video.
It’s that it doesn’t have to.
Instead of processing every frame equally:
Question → search → candidate moments → inspect deeply → answer.
AI is starting to reason about computation.
1
208
Romee Panchal retweeted
A useful rule in AI engineering:
Don’t start with the model. Start with the failure mode.
Ask:
• What can go wrong?
• Can rules solve part of it?
• Where does ML add value?
• How will we know it works?
Good AI starts with good problem framing.


