RLM: Recursive Language Models for Large Codebases - Shashi, Superagentic AI

What was stored

cues

135

whisperx 135

chunks

32

from 135 cues

keyframes

96

kept of 131 captured

frames with text

94

2,811 lines read

chapters

0

from the source metadata

keyframe bytes

18.0 MB

word timings on 135 cues

Provenance

Frames, and what the machine read

• 

  0:23 #0 done24 line(s)

  shot 0·sharpness 1606.4

  1. localhost1.00
  2. +0.90
  3. AI ENGINEER WORLD'S FAIR 20260.99
  4. Online Track1.00
  5. RLM:Recursive1.00
  6. LanguageModels1.00
  7. MIT RLM PAPER1.00
  8. Recursive Language Models1.00
  9. for Large0.95
  10. RLMs use a programmable environment for long-1.00
  11. context inference.1.00
  12. Codebases I0.91
  13. arXiv:2512.246010.99
  14. A coding-agent harness pattern, inspired by the MIT1.00
  15. RLM paper.0.99
  16. MIT RESEARCH1.00
  17. The RLM concept1.00
  18. MIT RLM paper: arXiv:2512.246010.99
  19. Use the RLM control loop as a coding-agent0.98
  20. harness pattern for large repository investigation1.00
  21. Shashi Jagtap1.00
  22. Founder, Superagentic Al1.00
  23. RLM: Coding-Agent Harness Pattern for Large Codebases0.99
  24. next back F fullscreen0.93
• 

  0:32 #1 done35 line(s)

  shot 1·sharpness 4950.1

  1. pecoming an independent0.97
  2. We gratefully acknowledge support from the Simons1.00
  3. Donate1.00
  4. onprofit.1.00
  5. Foundation, member institutions, and all contributors.0.99
  6. Search...1.00
  7. All fields1.00
  8. Search1.00
  9. Help | Advanced Search0.98
  10. Access Paper:1.00
  11. View P1.00
  12. HTML (e0.99
  13. nental)0.99
  14. TeX Source0.99
  15. (cc) BY0.93
  16. view license1.00
  17. Current browse context:0.99
  18. ns of inference-time scaling. We propose Recursive Language0.99
  19. cs.Al1.00
  20. nment and allows the LLM to programmatically examine,0.98
  21. < prev0.92
  22. next >0.98
  23. lly process inputs up to two orders of magnitude beyond model1.00
  24. new | recent | 2025-120.94
  25. ier LLMs and common long-context and coding scaffolds (e.g., on1.00
  26. Change to browse by:0.98
  27. deAct with sub-calls, and 13%against Claude Code) across four0.98
  28. CS0.66
  29. odel around the RLM. Our model, RLM-Qwen3-8B, outperforms0.99
  30. cs.CL1.00
  31. PT-5 on three long-context tasks. Code is available at this https0.98
  32. References & Citat0.99
  33. NASA ADS1.00
  34. Google Scholar1.00
  35. Semantic Scholar0.99
• 

  0:34 #2 done14 line(s)

  shot 2·sharpness 2010.0

  1. arxiv.org1.00
  2. C0.63
  3. Agent Harness Secret for Massive Codebases1.00
  4. https://arxi1.00
  5. Recursive Language Models1.00
  6. Alex L. Zhang0.99
  7. Tim Kraska1.00
  8. Omar Kh1.00
  9. MIT CSAIL0.95
  10. MIT CSAIL0.96
  11. MIT CSA0.98
  12. [email protected]1.00
  13. [email protected]1.00
  14. [email protected]1.00
• 

  0:37 #3 done19 line(s)

  shot 3·sharpness 1818.1

  1. LM:Recursive1.00
  2. MIT RLM PAPER1.00
  3. anguage Models0.97
  4. Recursive L0.99
  5. or Large0.98
  6. RLMs use a progra1.00
  7. context inference.1.00
  8. odebases lı0.89
  9. arXiv:2512.246011.00
  10. oding-agent harness pattern, inspired by the MIT1.00
  11. A paper.0.89
  12. MIT RESEARCH1.00
  13. The RLM co0.95
  14. T RLM paper: arXiv:2512.246010.99
  15. Use the RI M contr0.99
  16. ha1.00
  17. fo1.00
  18. ashi Jagtap1.00
  19. under, Superagentic Al0.98
• 

  0:41 #4 skipped

  shot 4·duplicate of #0

• 

  0:46 #5 done22 line(s)

  shot 5·sharpness 1231.0

  1. <0.88
  2. localhost1.00
  3. +0.87
  4. The problem1.00
  5. Large repositories require stateful, structured investigation.1.00
  6. SMALL REPO1.00
  7. LARGE REPO0.99
  8. MORE CONTEXT0.99
  9. REAL PROBLEM1.00
  10. One file, one test, one0.99
  11. Lost architecture,1.00
  12. Bigger windows add1.00
  13. A codebase is a1.00
  14. function. Agents handle1.00
  15. ownership, and cross-file0.99
  16. noise. Tool output floods1.00
  17. structured system, not a0.99
  18. this well.1.00
  19. links.1.00
  20. the chat.1.00
  21. long document.0.99
  22. RLM: Coding-Agent Harness Pattern for Large Codebases1.00
• 

  1:17 #6 done25 line(s)

  shot 6·sharpness 1311.0

  1. <0.92
  2. localhost1.00
  3. +0.89
  4. How people tackle it today0.98
  5. Search, retrieval, and longer context all help.1.00
  6. FILESYSTEM TOOLS1.00
  7. SEMANTIC RETRIEVAL1.00
  8. LONG CONTEXT1.00
  9. AGENT MEMORY0.97
  10. Search the repo directly0.99
  11. Find likely-relevant0.99
  12. Load or compress0.97
  13. Persist working state0.99
  14. grep, file readers, virtual FS.0.99
  15. code1.00
  16. evidence1.00
  17. Scratchpads and task state1.00
  18. Good when paths are known.1.00
  19. Embeddings narrow the0.98
  20. Bigger windows help, but detail1.00
  21. avoid rereading every token.0.99
  22. haystack, but can miss1.00
  23. competes with noise.1.00
  24. structure.1.00
  25. RLM: Coding-Agent Harness Pattern for Large Codebases0.99
• 

  1:58 #7 skipped

  shot 7·duplicate of #6

• 

  2:08 #8 done8 line(s)

  shot 8·sharpness 1383.2

  1. ontext management1.00
  2. 1. The repository is treated0.99
  3. to a programmable0.99
  4. 2. The model writes code to1.00
  5. xecution1.00
  6. nvironment.1.00
  7. 3. Focused subquestions re1.00
  8. rness Pattern for Large Codebases1.00
• 

  2:35 #9 done12 line(s)

  shot 9·sharpness 1272.6

  1. localhost1.00
  2. THE THESIS0.98
  3. RLMexternalizes1.00
  4. MIT RLM paper: arXiv:2512.246010.97
  5. context management0.98
  6. 1. The repository is treated as data the model can operate on.0.99
  7. into aprogrammable0.97
  8. 2. The model writes code to inspect, slice, and compute.0.99
  9. execution1.00
  10. environment.1.00
  11. 3. Focused subquestions return clean values.1.00
  12. RLM: Coding-Agent Harness Pattern for Large Codebases0.99
• 

  3:00 #10 skipped

  shot 10·duplicate of #9

• 

  3:31 #11 done35 line(s)

  shot 11·sharpness 1355.5

  1. <0.87
  2. localhost1.00
  3. +0.95
  4. 口0.60
  5. The analogy0.98
  6. A senior engineer with a notebook.0.98
  7. Q1.00
  8. Large project1.00
  9. Files, docs, tests, configs0.99
  10. Focused helper1.00
  11. task1.00
  12. inspect this1.00
  13. Notebook1.00
  14. Persistent REPL state1.00
  15. Engineer doing1.00
  16. evidence1.00
  17. the1.00
  18. Investigation scripts1.00
  19. Model-written Python0.99
  20. investigation1.00
  21. evidence, and open1.00
  22. keeps notes,1.00
  23. questions1.00
  24. Clean note1.00
  25. Focused helper1.00
  26. question1.00
  27. Ilm_query(...)0.94
  28. returned1.00
  29. added back to the0.98
  30. Returned note1.00
  31. A clean value for synthesis1.00
  32. notebook1.00
  33. Recursion = the parent REPL calling 11m_query(. .. )0.98
  34. antting a clean value back.0.99
  35. RLM: Coding-Agent Harness Pattern for Large Codebases0.99
• 

  3:40 #12 skipped

  shot 12·duplicate of #11

• 

  4:05 #13 skipped

  shot 13·duplicate of #11

• 

  4:42 #14 skipped

  shot 14·duplicate of #11

• 

  5:08 #15 done17 line(s)

  shot 15·sharpness 883.4

  1. localhost1.00
  2. +0.92
  3. The loop0.95
  4. How the harness actuallyruns.1.00
  5. Repo context1.00
  6. Root model1.00
  7. Python REPL1.00
  8. context1.00
  9. writes repl code1.00
  10. bounded observation1.00
  11. Recursive subcall0.99
  12. Clean result1.00
  13. Finish1.00
  14. llm_query(...)0.97
  15. value returns1.00
  16. FINAL / FINAL_VAR0.99
  17. RLM: Coding-Agent Harness Pattern for Large Codebases0.98
• 

  5:38 #16 done24 line(s)

  shot 16·sharpness 1199.5

  1. <0.86
  2. localhost1.00
  3. +0.94
  4. 口0.57
  5. Why codebases0.98
  6. Repos are structured and executable.0.99
  7. REPOSITORY AS STRUCTURED DATA1.00
  8. runner.py1.00
  9. orchestration1.00
  10. environment.py1.00
  11. REPL state0.97
  12. 1 Files and directories give the model a map.0.98
  13. 2 Imports and symbols reveal execution paths.1.00
  14. termination.py1.00
  15. FINAL /FINAL_VAR0.97
  16. Tests and fixtures expose intended behavior.0.99
  17. tests/1.00
  18. behavior checks1.00
  19. 4 Configs and docs explain runtime assumptions.0.99
  20. README.md1.00
  21. usage contract1.00
  22. MODEL-WRIEN INSPECTION0.99
  23. extract ewnce → llm_query → FINAL0.95
  24. RLM: Coding-Agent Harness Pattern for Large Codebases0.99
• 

  5:53 #17 skipped

  shot 17·duplicate of #16

• 

  6:15 #18 done10 line(s)

  shot 18·sharpness 818.4

  1. localhost1.00
  2. è0.72
  3. +0.77
  4. 020.93
  5. RLMCode1.00
  6. An independent experimental reference implementation used to demonstrate the pattern.0.99
  7. Landing page1.00
  8. Docs1.00
  9. GitHub1.00
  10. RLM: Coding-Agent Harness Pattern for Large Codebases1.00
• 

  6:18 #19 done6 line(s)

  shot 19·sharpness 502.0

  1. 020.98
  2. RLMCode1.00
  3. An independent experimental reference implementation used to der1.00
  4. Landing1.00
  5. Docs1.00
  6. GitHub1.00
• 

  6:22 #20 done31 line(s)

  shot 20·sharpness 1892.3

  1. super-agentic.ai1.00
  2. X0.67
  3. +0.94
  4. RLM: The Coding Agent Harness Secret for Massive Codebases0.98
  5. RLM Code | Research Playground for Recursive Language Models | Superagentic Al1.00
  6. SuperagenticAI1.00
  7. <> Products0.99
  8. Solutions1.00
  9. OSS0.76
  10. Research1.00
  11. Resources0.95
  12. Company1.00
  13. RESEARCH-FIRST EVALUATION OS1.00
  14. RLMCode1.00
  15. Research Playground for Recursive0.99
  16. Language Models1.00
  17. Run real RLM workflows, benchmark them, replay every step, and1.00
  18. compare results under controlled budgets and secure sandboxes.1.00
  19. Open source under Apache-2.0. Explore the project on GitHub.0.98
  20. Start with Docs0.99
  21. View GitHub1.00
  22. Apache-2.01.00
  23. PyPI Package1.00
  24. Research-first1.00
  25. What is included1.00
  26. Preferred stack1.00
  27. Code Mode support1.00
  28. Watch demo1.00
  29. Quick workflow1.00
  30. Security1.00
  31. FAQ1.00
• 

  6:23 #21 done32 line(s)

  shot 21·sharpness 1924.5

  1. super-agentic.ai1.00
  2. X0.60
  3. +0.95
  4. RLM: The Coding Agent Harness Secret for Massive Codebases1.00
  5. RLM Code | Research Playground for Recursive Language Models | Superagentic Al0.99
  6. SuperagenticAI1.00
  7. <> Products0.99
  8. Solutions1.00
  9. OSS0.72
  10. Research1.00
  11. Resources0.95
  12. V0.82
  13. Company0.91
  14. RESEARCH-FIRST EVALUATION OS0.99
  15. RLMCode1.00
  16. Research Playground for Recursive1.00
  17. Language Models1.00
  18. Run real RLM workflows, benchmark them, replay every step, and1.00
  19. compare results under controlled budgets and secure sandboxes.1.00
  20. Open source under Apache-2.0. Explore the project on GitHub.0.98
  21. Start with Docs0.99
  22. View GitHub1.00
  23. Apache-2.01.00
  24. PyPI Package1.00
  25. Research-first1.00
  26. What is included0.99
  27. Preferred stack1.00
  28. Code Mode support1.00
  29. Watch demo1.00
  30. Quick workflow1.00
  31. Security1.00
  32. FAQ1.00
• 

  6:27 #22 done51 line(s)

  shot 22·sharpness 2263.7

  1. super-agentic.ai1.00
  2. ×0.79
  3. +0.97
  4. RLM: The Coding Agent Harness Secret for Massive Codebases1.00
  5. RLM Code | Research Playground for Recursive Language Models | Superagentic Al0.99
  6. SuperagenticAI1.00
  7. <> Products0.95
  8. Solutions1.00
  9. Oss0.54
  10. Research0.95
  11. Resources0.97
  12. Company0.97
  13. V0.90
  14. RESEARCH-FIRST EVALUATION OS1.00
  15. RLM Code0.97
  16. RLM1.00
  17. RLM Code1.00
  18. Research Playground for Recursive1.00
  19. CODE-0.98
  20. Research Playground & Evaluation OS for RLM Agentic Systems1.00
  21. Language Models1.00
  22. RLM0.86
  23. Files0.95
  24. Details0.97
  25. Shell1.00
  26. Research Lab0.97
  27. One Screen:0.97
  28. Quit0.80
  29. Run real RLM workflows, benchmark them, replay every step, and1.00
  30. RLM Restarch Lib0.83
  31. compare results under controlled budgets and secure sandboxes.0.99
  32. Dashboard1.00
  33. Trajectory1.00
  34. Benchnarks0.99
  35. Replay0.99
  36. Events0.99
  37. Open source under Apache-2.0. Explore the project on GitHub.0.99
  38. Rum Dashboard0.98
  39. Start with Docs1.00
  40. View GitHub1.00
  41. aolong_style0.94
  42. Apache-2.01.00
  43. PyPI Package1.00
  44. Research-first1.00
  45. What is included0.99
  46. Preferred stack1.00
  47. Code Mode support1.00
  48. Watch demo1.00
  49. Quick workflow1.00
  50. Security1.00
  51. FAQ1.00
• 

  6:28 #23 done11 line(s)

  shot 23·sharpness 694.4

  1. RLM: The Coding Agent Harness Secret fic0.97
  2. SuperagenticAI1.00
  3. <> Products0.97
  4. 目0.67
  5. Solutions0.94
  6. 0SS0.60
  7. RESEARCH-FIRST EVALUATION OS0.99
  8. RLMCode1.00
  9. Research Playground for Recursive1.00
  10. Language Models0.97
  11. Pun real Rl M workflowe. henehmark them mwvay every st ay0.83

Transcript

135 cues· 2,394 words· 12,806 chars