Supreeth Ravi
7 posts
Supreeth Ravi
@supreeth___ravi
Building the AI infrastructure of the future. Head of Eng & Product @PhroneticAI. PayCentral -India's first agentic payment platform. TedX Speaker, Founder
San Francisco Katılım Ekim 2014
61 Takip Edilen5 Takipçiler
Hey @hwchase17 fair points, LangGraph + smart tool patterns get you super close already.
Files as vars, sub-agents, dumping intermediates… it’s like 85% of the RLM vibe in practice.
But the stuff @lateinteraction keeps stressing really shines on long/nested tasks :
1. treating the whole user prompt as an untouchable symbolic object (no greedy slurping big chunks into context)
2. true deep recursion inside code execution loops without bloating the main prompt
3. strict no-pollution everything stays in vars until you decide otherwise
That combo is what let PolyRAG chain ColBERT + GraphRAG + memory lookups without context exploding or re-embedding everything constantly. Feels noticeably cleaner on hairy multi-hop stuff.
Curious tho have you guys in LangChain land tried forcing agents into that exact “prompt-as-immutable-var + deep-in-loop recursion + zero message leakage” mode?
Any configs that get weirdly close?
Check out the full PolyRAG writeup if you’re curious how I leaned into it: open.substack.com/pub/supreethra…
English
good thread on whether RLMs are just coding agents or not
I am still in the camp they are basically just coding agents with some distinct features (just like some coding agents have different subagent/swarm features)
eg @lateinteraction in response your points, isn't that just coding agents with:
1. files as input (can be done with UX, eg in claude when you post long snippets it makes it a file at least in ui)
2. just give coding agent a cli command that itself/a subagent
3. with (2) done it can trivially save it to files
Omar am i missing anything?
Omar Khattab@lateinteraction
The following are not standard in a coding agent: 1. The user prompt P itself (not just external data) is a symbolic object in the environment. The model is not allowed to grep/read long snippets from P. 2. The model has to write recursive code (that calls LMs) to understand or transform the content of P. Unlike "sub-agents", recursion must happen during code execution, which means that you can launch arbitrarily many sub-calls, not just a small constant number, without polluting the context window. 3. All sub-calls and tool calls return values into symbolic variables. The model is not allowed to pollute its context window with their return values. Instead, it must build up (and refine) its output with recursion. Contrast Algorithm 1 and Algorithm 2 to see the formal differences.
English
Supreeth Ravi retweetledi
Can probabilistic AI systems safely execute deterministic financial transactions?
We ran an empirical study across 160,000 simulated payments and real LLM API validations.
The result: direct LLM → payment execution introduces systemic, unfixable failures — regardless of model quality.
This research, authored by @supreeth___ravi , also proposes a deterministic mandate-based architecture that eliminates these failures by design.
Research summary ↓
phronetic.ai/whitepaper
#AI #LLMs #FinTech #Payments #Systems
English
Supreeth Ravi retweetledi
🔍 Agent Skills are quickly becoming an open standard across the AI ecosystem.
Anthropic introduced it.
OpenAI adopted it days later.
The specification is clear. Implementation is not.
This article, authored by @supreeth___ravi , examines the gap between specification and production, the role of progressive disclosure, and what it takes to make Agent Skills usable in real-world agent systems.
phronetic.ai/blogs/anthropi…
#AgenticAI #AIInfrastructure #DeveloperTools
English
Your Heart is safe with US..! supreethsln.wordpress.com/2016/07/27/you… via @wordpressdotcom
English