Interactive Explainer

Agent Harness Anatomy

The model is only half the agent. The harness — context delivery, tool interfaces, planning artifacts, memory, guardrails, and evaluation loops — is the scaffolding that determines whether an AI agent succeeds or fails on real tasks. Explore each system below, then simulate how harness quality changes outcomes.

Context deliveryTool interfacesPlanning artifactsMemoryGuardrailsEval loops

The Six Systems of a Harness

Click or tab through any component in the diagram. Requests flow down from the task; observations flow back up through the loop.

Real taskContext deliverywhat the model seesPlanning artifactsplans, todos, subgoalsMemory & statescratchpads, summariesTool interfacesactions on the worldModel corereasoning inside the loopGuardrailspermissions, budgets, checksEvaluation loopsdid it actually work?safe outputsverified result

Harness Quality Simulator

Same model, different harness. Set the maturity of each system, then run a simulated 12-step task and watch where the run breaks down. Scores use a simple compounding model: each step must survive every subsystem.

--task success

Adjust the sliders and press Run simulated task to see how far the agent gets.

No run yet. The trace of each simulated step will appear here.

Design Primitives

Recurring primitives from harness engineering practice — small ideas that compound across long tasks.

Fresh context per step

Rebuild the prompt from durable state instead of letting a raw transcript grow. Stale context is the quietest failure mode in long-horizon agents.

Legible tools

Few, well-named tools with strict schemas and rich error messages beat many overlapping ones. The error message is part of the interface.

Externalized plans

Plans, todo lists, and subgoal files live outside the model so progress survives context resets and can be inspected by humans.

Verification over trust

Never accept the agent's claim of success. Run tests, diff outputs, and check invariants after every meaningful action.

Budgets everywhere

Cap steps, tokens, retries, and spend. A harness that cannot stop an agent cannot be trusted to run one.

Recoverable failure

Design for retry: idempotent actions, checkpoints, and summaries so a failed step costs one step, not the whole run.

Harness Design Checklist

Score your own agent scaffold. Check what you have, then copy the checklist to share or file as an issue.

0 / 12 covered

Harness anatomy: multiplicative assumptions and simulated traces

Read the explanation

The saved harness explainer contains six maturity sliders: context, tools, planning, memory, guardrails and evaluation. Each becomes an assigned factor of zero point six two plus zero point three eight times maturity divided by one hundred. The simulator multiplies all six factors to obtain one per-step probability, capped at zero point nine nine five. With every slider sixty, each factor is zero point eight four eight and their product is about zero point three seven one eight five six. With every slider ninety five the product is about zero point eight nine one two eight zero. Bars compare these products at three hundred pixels per assigned probability. This is a mathematical assumption in the page, not a measured relationship between harness maturity and actual model reliability. The category labels explain design ideas, but no benchmark or model invocation validates the multipliers. The displayed expected task success is the per-step probability raised to twelve, because the saved trace has twelve named steps and stops at the first failure. As a mathematical illustration, a per-step probability of point nine gives about point two eight two four three zero for a twelve-step run, while point nine nine gives about point eight eight six three eight five. Bars show about eighty four point seven three and two hundred sixty five point nine two pixels at three hundred pixels per full-run probability. This assumes independent trials with the same probability; correlated failures or changing task difficulty are outside the model. At the default slider value sixty, the page assigns expected success around point zero zero zero six nine nine percent, while ninety five assigns about twenty five point one two nine percent. Those extreme values result from repeating six multipliers across twelve steps, not observed agent performance. At maturity one hundred for every category, the uncapped product is one, but stepProb limits it to point nine nine five. Across twelve steps, expected success is about point nine four one six two three, or ninety four point one six two percent. Bars compare the capped step probability and full-run probability at three hundred pixels per probability unit. A click draws independent Math.random values and produces a local trace; a successful trace does not change that expected gauge value. On failure, the named broken system is whichever slider is lowest, rather than a causal diagnosis of the random event. On equal sliders the first category stays selected as weakest. Moving sliders or using the strong and weak presets does not recompute the gauge until Run is clicked. Reset restores sixty and clears the previous score. The twelve checklist items are counted and copied, but checking them does not affect simulation probability. This explainer describes saved educational logic without claiming a real agent completed the trace.

Copied to clipboard
Super generates helpful tools and automates fact-checking across the internet proactively. If you enjoyed this tool, build your own with Super and share it with a friend.