Don't put the repo into context — let the model write code to inspect it and recursively call sub-models; the open source RLM implementation, demoed live end to end
RLM: Recursive Language Models for Large Codebases - Shashi, Superagentic AI · Shashi
17 min total·Actually worth watching closely: ~4 min·3 must-watch clips
- 0:00 – 1:30Skim
Opening: RLM and the MIT paper
Introduces the speaker and the topic, points to the RLM (recursive language models) paper published by MIT and friends, and frames the talk as going from the concept through to an open source implementation and a demo.
There is a full paper you can read; this talk is about using its concepts in your own workflow and harnesses
Mostly the paper page and title slides on screen (29.96s, 'as you can see, there is a full paper') — a glance to note where it comes from is enough▶ Jump to 0:00Speaker · Shashi - 1:30 – 4:10Listen
Context degradation and the four existing approaches
Coding agents work well on small repos but degrade as context grows, and it gets worst on large mono repos; existing approaches fall into four buckets: grep-style search, semantic or local search, compressing long context into a summarized version, and bolt-on memory.
All four mainstream approaches are about putting better content into the context, while RLM changes the approach entirely
Purely spoken problem framing and taxonomy, no key visuals — fine to listen to like a podcast▶ Jump to 1:30Speaker · Shashi - 4:10 – 7:01Skim
How RLM works
RLM externalizes context management into a programmable execution environment: the whole repo is data the model can operate on, the model writes REPL code to slice and compute and gets back a bounded observation, and when the loop needs more information it goes through the LLM query to recursively ask another model, looping until it reaches a final result.
The core is bounded observation plus recursive sub-calls, replacing everything-into-context
The mechanism is explained over architecture slides, and at 361s the fully open source RLM Code repo is announced — reading the loop diagram is faster than listening alone▶ Jump to 4:10Speaker · Shashi - 7:02 – 9:06Listen
The RLM ecosystem: a pattern, not a framework
RLM is a pattern you can implement in your own way: the official repos are RLM and RLM Minimal, various other people have their own implementations, and Omar — RLM's author — also authored dspy, which has RLM inside it, though the two should be treated as completely different.
Learn RLM as a pattern rather than binding yourself to one specific framework
Mostly spoken narrative about the ecosystem and the people behind it; the slides carry little information▶ Jump to 7:02Speaker · Shashi - 9:06 – 10:46Skim
RLM Code and the demo setup
Introduces Super Agent's open source RLM Code: the paper implemented as it is with nothing added on top, running with local or cloud based models and pluggable into any observability framework; the demo uses Docker as a sandbox, with RLM Code's own source as the demo target, and the companion repo is there to reproduce it yourself.
The whole demo is reproducible: the RLM paper, RLM Code and the demo repo are all open
546s shows the companion repo's readme and 605s the demo target config — note the repo entry points and leave the details for when you read the code▶ Jump to 9:06Speaker · Shashi - 10:46 – 12:44Watch
Live demo: the full RLM loop
A real run inside the Docker sandbox: first the model writes the REPL code and builds the evidence, then it makes the LLM query call with a prompt and gets the result back, then it reaches the final answer; from 687s the trace details, and at 728s the runs and sessions in JSONL that plug into any observability platform.
The RLM loop really runs, and every step leaves a trace you can inspect
The core visuals of the talk: the REPL code, the recursive call and the token counts unfolding in the terminal only make sense on screen — 654s, 687s and 728s are all flagged as key visual moments▶ Jump to 10:46Speaker · Shashi - 12:44 – 15:00Watch
The TUI harness and the research lab
Demos the coding agent style experimental TUI: connect with the Gemini model, run doctor to check things, and specify a budget when sending the prompt so the run doesn't spend too much, under maximum steps and recursion depth (801s); from 848s, the research lab replays the rewards and the whole trajectory.
A homegrown RLM harness needs budget, step and recursion depth limits plus full trajectory capture
The TUI flow and the research lab's trajectory replay are purely visual; the three visual moments at 764s, 801s and 848s cluster in this section▶ Jump to 12:44Speaker · Shashi - 15:00 – 17:26Listen
Real-world use cases and industry adoption
The natural fits are root cause analysis on large source code and onboarding onto unfamiliar repos; the Codex harness was seen writing Python in the REPL to curate context, dynamic workflows where multiple agents each get their own sandbox generally come from RLM, and Claude Code engineers have acknowledged on X that they used RLM concepts.
The RLM pattern is already inside mainstream products like Codex and Claude Code — reason enough to design a harness for your own case now
The closing is mostly spoken industry trends and advice; 899s only points at the available source code, so there's no need to watch▶ Jump to 15:00Speaker · Shashi