Agent Scaffolding
ELI5 - The Vibe Check
Agent scaffolding is the training wheels and guardrails wrapped around a model so it acts like a competent worker rather than a genius with no discipline. It's the prompts, retry logic, and structured steps that turn 'brilliant but occasionally unhinged' into 'reliably finishes the task.' Better scaffolding can make a smaller model outperform a bigger one that's left to freestyle.
Real Talk
Agent scaffolding refers to the prompting, tool schemas, control flow, and error-handling code layered around a model to make it perform agentic tasks reliably rather than relying on the raw model's judgment at every step. On coding benchmarks like SWE-bench, improvements to scaffolding alone have produced larger score jumps than swapping in a stronger underlying model.
When You'll Hear This
"Better scaffolding got us a 15-point jump on the benchmark with the same model." / "We rewrote the scaffolding, not the prompt."
Related Terms
Agent Harness
The agent harness is the exoskeleton around a raw model. On its own, an LLM is just a very smart text predictor sitting in a box.
Guardrails
Guardrails are the safety nets you put around AI applications — rules and checks that prevent the AI from going rogue.
Prompt Engineering
Prompt engineering is the art of talking to AI so it actually does what you want.
Tool Use
Tool use is when an AI can call external functions, APIs, or programs to do things it can't do alone.