AI + Data guy @SentientAGI | prev-@GnaniAi, @deloitte
India
Joined June 2019
- Tweets754
- Following460
- Followers64
- Likes1.4K
Pinned Tweet
@omarsar0 Thank you very much for sharing my work!
check it out: arxiv.org/abs/2607.22520
Are agent skills always worth using?
The answer is no?
This paper provides some important insights to understand this more.
(bookmark it)
Paper summary:
Adding procedural skills to an agent is usually scored by average task success. That number nets gains against damage and hides half of what happened.
Setup: nearly 6,000 paired runs across two office automation benchmarks and three model harness stacks, comparing the same agent with and without skills.
A regression is a task the agent solved without skills and then failed once skills were added. Regressions are large enough that the best performing skills separate themselves mainly through fewer regressions. Larger gains contribute much less.
Three mechanisms drive it.
Skill description osmosis, where a skill changes agent behavior just by sitting in context even when it is never invoked. Grounding displacement, where a prescribed procedure overrides how the agent reads its inputs. Verification displacement, where the procedure suppresses checks the agent would otherwise run on its own output.
Trace analysis surfaces that procedural guidance is the stage least often responsible for failure, while grounding and verification dominate the errors that remain. Existing skills are almost entirely procedure.
Paper: arxiv.org/abs/2607.22520
Learn to build effective AI agents in our academy: academy.dair.ai/
General observation: Many research papers from top universities are project first and that's the actual way of publishing papers rather than going random
Darshan Tank retweeted
This new paper is interesting in that it more rigorously confirms something we’ve all seen anecdotally—AI agents get distracted quite easily.
Both skill design / depth, appropriate for the intelligence of the model, and skill injection methods, are critical to ensure the LLMs don’t regress severely when used inside complex reasoning harnesses.
Are agent skills always worth using?
The answer is no?
This paper provides some important insights to understand this more.
(bookmark it)
Paper summary:
Adding procedural skills to an agent is usually scored by average task success. That number nets gains against damage and hides half of what happened.
Setup: nearly 6,000 paired runs across two office automation benchmarks and three model harness stacks, comparing the same agent with and without skills.
A regression is a task the agent solved without skills and then failed once skills were added. Regressions are large enough that the best performing skills separate themselves mainly through fewer regressions. Larger gains contribute much less.
Three mechanisms drive it.
Skill description osmosis, where a skill changes agent behavior just by sitting in context even when it is never invoked. Grounding displacement, where a prescribed procedure overrides how the agent reads its inputs. Verification displacement, where the procedure suppresses checks the agent would otherwise run on its own output.
Trace analysis surfaces that procedural guidance is the stage least often responsible for failure, while grounding and verification dominate the errors that remain. Existing skills are almost entirely procedure.
Paper: arxiv.org/abs/2607.22520
Learn to build effective AI agents in our academy: academy.dair.ai/
Do we have harness specifically for code-review that is token efficient and upto the mark?
Doing review with Astra kills usage limit half way!
Just a thought: A new benchmark should measure not only how many tasks a model can solve, but also how many tokens it consumes to complete each task.
I've enjoyed working with Cursor pre-acquisition and have respect for the team and what they have built. The 5% here should have come with strong caveat and I would love for Michael to share the math.
Tokens are not a proxy for revenue nor value created and the OpenAI models are on the very frontier of token efficiency. Smaller or less strong models require many more tokens to achieve a task and therefore will inflate traffic share significantly.
We got gta 6 before gta 6
ALT Dear all, We know that many of you have been waiting to hear from us regarding the events of the past week. It would be an understatement to say that having videos of Grand Theft Auto VI gameplay leak in this way has been heartbreaking for our team, and this is obviously not how we intended for you to see the game after all this time. We are very sorry that everything has taken as long as it has \u2014 from getting the game finished (nearly there!) to sharing more details and official gameplay, and providing the community with everything you want to know. We are very excited for everyone to see the extended look tomorrow. It has taken longer than we wanted to get it ready, but as with everything we do, we know we need to exceed your expectations, and we are determined to deliver at the level you expect and deserve.
I don’t understand why FDE is so hyped up!
It’s basically client facing role where you help your client integrate and use AI!
Same shit we used to do at gnani ai back in 2023-2024!
Darshan Tank retweeted
“CryptoAnalystBench: Failures in Multi-Tool Long-Form LLM Analysis” has been accepted as a poster at KDD 2026 (@kdd_news) — one of the leading conferences for AI, machine learning, and data mining.
At KDD? Meet with Sentient researcher Darshan Tank (@TankDarshan7) to learn why LLM agents struggle with multi-tool, long-form analysis in crypto research.
