Glossary · Term

scaffold

← all terms

Definition

Plain language

The wrapper of code, tools, and loops around an AI model that turns the raw model into an agent that can actually do things.

As stated in the literature

The surrounding software that drives a stateless model in an action-observation loop — context assembly, tool dispatch, stopping logic, sandboxing, logging; can change the same model's token usage and behavior by an order of magnitude, so model comparisons must hold it fixed.

Also called: scaffolds, scaffolding

Why it matters: It can change a model's behavior and cost by a huge margin, so fair comparisons between models must keep the scaffold fixed or risk measuring the wrapper instead of the model.

For example, the same model can either flail or solve a coding task depending on the surrounding code that feeds it context, runs its tools, and decides when it's done.

Heard on the show

“The scaffolding you never think to inspect.”
Episode 210 — Same Website Request, Different Code — The Bias You Can't See

Mentioned in 67 episodes

  1. 210
    Same Website Request, Different Code — The Bias You Can't See
  2. 200
    The One Mechanism That Turns Twenty AI Clones Into an Actual Team
  3. 192
    A 32B Open Model Matched Frontier Systems By Learning to Take Notes
  4. 191
    How One Researcher Beat GPT-5.2 and Gemini 3 by Judging Their Answers, Not Improving Them
  5. 184
    An AI Built an Undetectable Secret Channel, And Another AI Couldn't Find It
  6. 175
    One Crosscoder Feature Flips a Stalling Chatbot Into a Working Agent
  7. 166
    A Router That Beats the Frontier Models It Calls
  8. 163
    Why Training Only on Perfect Solutions Cripples a Model's Reasoning
  9. 160
    Training an AI to Take Its Own Notes, So Its Future Self Works Better
  10. 157
    When an AI Coding Agent Drives a Phone Through the Terminal, No Screen Needed
  11. 155
    Why a Flawless Demo Makes a Worse Computer-Using Agent, And the Fix
  12. 152
    Training a Model to Mean What It Says, And Why That Isn't the Same as Being Good
  13. 151
    Why More Experience Made This AI Agent Worse, And How to Fix It
  14. 150
    Don't Kill the Loser: A Different Way to Handle Two AI Agents Colliding
  15. 147
    Agents Fail at the Body, Not the Brain: A Self-Rewriting Scaffold That Lifts a 9B Model 44 Points
  16. 146
    How an Innocent README Can Freeze an AI Agent's Safety Check for an Hour
  17. 144
    When an AI Agent Just Copies Its Tool — And Bigger Models Copy More
  18. 142
    Training a Tiny Model to Run the Plumbing Between an Agent and the World
  19. 140
    When a Reasoning Model Says "Let Me Double-Check" After It's Already Decided
  20. 139
    When Optimizing One GPU Kernel Quietly Breaks the Whole System
  21. 133
    How MiniMax Turned a Reward-Hacking Disaster Into Olympiad Gold
  22. 131
    Why Autonomous Research Agents Forget Their Own Lessons, and Arbor's Fix
  23. 130
    Why AI Agents Coordinate Better Through a Shared Board Than a Boss
  24. 129
    How a Crowd of Anonymous AI Agents Broke a 40-Year Math Record
  25. 125
    AI Coding Agents Run a Marathon, and Fewer Than One in Three Finish
  26. 121
    When the Agent Says It's Done But Nothing Happened: Debugging the Harness, Not the Model
  27. 120
    How an AI Agent Rewrites Its Own Tools, Without an Answer Key
  28. 118
    Why the Best-Aligned AI Models Are the Easiest to Trick Into Producing Harm
  29. 114
    Agents That Rewrite Their Own Weights Instead of Just Taking Notes
  30. 112
    When an AI Agent Cheats Without Being Told: Inside the Meta-Agent Challenge
  31. 110
    How an Agent Got 44 Points Better by Mining Its Own Scratch Paper
  32. 109
    An AI Got Caught Reading the Answer Key, And Why That Catch Matters
  33. 106
    Giving Agents a Notebook Instead of New Weights: How ExpGraph Lets Frozen Models Learn
  34. 102
    How to Catch an AI Attack That No Single Conversation Reveals
  35. 101
    Treating Math Formalization Like a Codebase, and Where the Agents Cheat
  36. 100
    How a Prompt Wrapper Lets a Frontier Model Play Poker Like an Expert
  37. 097
    Same Tokens, Same Cost, Wildly Different Results: What Actually Scales in AI Agents
  38. 090
    How MiniMax-M2 Bets That Sparsity Plus Verifiable Rewards Can Match Frontier Agents
  39. 088
    Two Levers for Self-Improving AI: When Rewriting Code Isn't Enough
  40. 079
    An Old Idea From Cognitive Psychology Reshapes How We Reward Reasoning Models
  41. 078
    Training a Markdown File: When LLM Self-Improvement Borrows the Discipline of Neural Net Training
  42. 077
    Reading a Model's Confidence Curve to Decide When Chain-of-Thought Is Worth It
  43. 076
    Same Model, Organized Differently: How an Agent Architecture Beat Frontier Systems at Research Math
  44. 075
    Growing Code and Proof Together: Verified Systems in Ten Hours Instead of a Year
  45. 073
    When Three LLMs Talk to Each Other, Their Ideas Quietly Stop Moving
  46. 072
    A Robot Made Graphene Without Help, And Caught Itself Hallucinating
  47. 071
    When the Model Is Fine and the Plumbing Is Broken: Fixing Agents at the Interface
  48. 067
    An AI Just Solved a 1996 Erdős Problem—and the Simplest Agent Won
  49. 065
    One Loop to Optimize Them All: A Universal API for LLM-Driven Discovery
  50. 060
    When Splitting One Model Across Three Agents Doubles Its Accuracy
  51. 053
    An AI Agent Swapped In Focal Loss And Beat A Human-Tuned Training Script
  52. 051
    Why Parallel Sampling Plateaus, And What Evidence Graphs Do Instead
  53. 047
    When Agent Benchmarks Lie: The Harness Problem in Open-Source AI
  54. 046
    When the AI Optimizer Edits the Grade Book: Why Harnessing Evolution Needs a Wall
  55. 041
    When the Iteration Teaches the Model to Skip the Iteration
  56. 036
    Sparse Attention Was the Wrong Frame. Treat It as Geometry Instead.
  57. 028
    Teaching a Model to Hire Copies of Itself: Recursive Agent Optimization
  58. 027
    When AI Agents Build the Serving Stack: A Bet on Bespoke Infrastructure
  59. 024
    An AI Agent That Found 28 Zero-Days in Windows — And What Made It Work
  60. 017
    When the Agent Grades Its Own Homework: A Brutal New Benchmark for AI Workers
  61. 014
    Why a Constrained Pipeline Beat a Full Coding Agent at Finding Bugs 30-to-1
  62. 013
    Why Search Keeps Rediscovering the Same Workflow, and What That Means
  63. 012
    Why AI Coding Agents Keep Trying to Debug Without a Debugger
  64. 008
    Why Long-Horizon AI Agents Get Stuck, and a Milestone-Based Fix That Helps
  65. 007
    Exploration Hacking: When Models Sabotage Their Own RL Training
  66. 002
    An AI Ran a Real Optics Lab for 21 Hours and Found a Transformer-Shaped Pattern in Light
  67. 001
    When AI Models Quietly Protect Each Other From Shutdown

Related concepts

Agent Benchmarks Agent Scaffolding Capability Elicitation Generation-Time Specialization Inference-Time Scaffolding Recursive Agent Optimization Subgoal Decomposition Workflow Search

Related terms

token