Skip to content
AI DemystifiedHow AI actually works
Lesson 17 of 22Advanced topics & capstoneadvanced10 min

Unhobbling AI

When the product is the bottleneck, not the model

What this lesson is about

Strong models often underperform because products hobble them. Unhobbling removes accidental blockers and adds intentional machinery: GTER loops (goal + tools + evaluator + retry), context packing, sandboxes, and selective tools. Inspired by practitioner talks such as Stop Hobbling Your AI. Teaching simulator only.

Guided mission · login 500

Step 1 / 7

The ticket

Production login returns 500. You have a coding agent, a test suite, and a product full of rails. Goal: ship a fix without babysitting every keystroke.

Effective capability
45
Product overhang
47
Context quality
0
Loop autonomy
35one-shot
Permissions: Confirm everything5 hobbles active

Effective capability will drop hard under product rails.

Related: Tools · Agents · Context · Prompts · Playground

Hobbling

The model is trying to work and the product (or workflow) keeps getting in the way: confirm dialogs, missing tools, truncated memory, human-only glue.

Product overhang

The model can already do more than the product exposes. The gap is capability you paid for but never elicited — today's model, not a future one.

Loops over prompts (GTER)

Goal + Tools + Evaluator + Retry. You design the loop once; the system re-prompts itself until green or budget.

Context engineering

What enters the window each step: write, select, compress, isolate. Wrong context hobbles even a strong model.

Parallel agents

One chat is a single worker. Fan out many agents on a backlog, then review exceptions — not every line.

Permissions ladder

Unhobbling is not removing safety. Least-privilege sandboxes beat confirm-everything micromanagement.

Harder problems, less hand-holding

As models improve, under-constrain: bigger goals, automated checks, budget shifted from repetitive human steps to tokens and evaluation.

Ablate scaffolding

Delete system-prompt scar tissue each model generation. Re-add lines only when failures keep repeating.

Core demo: product overhang

Flip rails. Effective capability is what the model delivers through the product.

Effective capability

51 / 100
Model can do92
Product lets through51
Product overhang41

Deep drills

Optional interactive labs. Expand what you want after the mission.

Practical takeaways

  • Audit confirmations, missing tools, human glue, and tool flooding.
  • Engineer context: write, select, compress, isolate.
  • Prefer GTER loops with hard evaluators plus soft rubrics.
  • Unhobble into sandboxes; HITL only on irreversible actions.
  • Ablate system prompts after model upgrades.

Honesty note

Numbers are didactic. Unhobbling removes accidental friction, not review, permissions, or evaluation. Practitioner productivity claims are domain-dependent hypotheses, not universal laws.

Check your understanding

1/7

What does “hobbling” mean in this lesson?

Sources, scope & review statusVerified Jul 29, 2026

Claims this lesson makes

  • Review frequentlyScaffolding, context assembly, tools, permissions, and evaluation can materially change observed agent performance.

    Scope: The GTER and overhang scores are explicit teaching models, not benchmark measurements.

Next content review: 2026-10-29. Source links establish the lesson's claims; interactive numbers remain labeled simulations unless explicitly identified as measured data.