Module decomposition cut agent token use 32% on follow-up feature additions

sgathrid1 pts1 comments

Agentic Cost Savings - Topos Documentation

Contents

Menu

Expand

Light mode

Dark mode

Auto light/dark, in light mode

Auto light/dark, in dark mode

Skip to content

Topos Documentation

Getting Started

Quick Start

Installation

Guides

For Agents

CLI Reference

Concepts

Measures

Concepts

Case Studies

Agentic Cost Savings

Reference

API Reference<br>Core<br>topos.core.category

topos.core.morphism

topos.core.object

topos.core.omega

Evaluation<br>topos.evaluation.characteristic_morphism

topos.evaluation.preferences

topos.evaluation.suggestions

topos.evaluation.suppression

topos.evaluation.policies.base

topos.evaluation.policies.calibration

topos.evaluation.policies.clones

topos.evaluation.policies.composable

topos.evaluation.policies.coverage

topos.evaluation.policies.secure

topos.evaluation.policies.simple

Graphs<br>topos.graphs.base

topos.graphs.ast.dispatch

topos.graphs.ast.object

topos.graphs.ast.types

topos.graphs.ast.providers.base

topos.graphs.ast.providers.native_provider

topos.graphs.ast.providers.tree_sitter_provider

topos.graphs.cfg.builder

topos.graphs.cfg.models

topos.graphs.cfg.object

topos.graphs.cpg.builder

topos.graphs.cpg.models

topos.graphs.cpg.object

topos.graphs.mdg.object

topos.graphs.pdg.object

topos.graphs.uast.mapper_common

topos.graphs.uast.mapper_cpp

topos.graphs.uast.mapper_javascript

topos.graphs.uast.mapper_python

topos.graphs.uast.mapper_rust

topos.graphs.uast.mapper_typescript

topos.graphs.uast.models

Functors<br>topos.functors.probes.ast.complexity

topos.functors.probes.ast.entropy

topos.functors.probes.cfg.complexity

topos.functors.probes.cfg.paths

topos.functors.probes.cpg.danger

topos.functors.probes.cpg.taint

topos.functors.probes.mdg.coupling

topos.functors.probes.mdg.fan

topos.functors.probes.uast.compare

topos.functors.probes.uast.signature

topos.functors.profunctors.ast.compare

topos.functors.profunctors.cfg.compare

topos.functors.profunctors.cpg.compare

topos.functors.profunctors.cpg.topological_coverage

topos.functors.profunctors.mdg.compare

topos.functors.profunctors.pdg.compare

topos.functors.profunctors.uast.compare

topos.functors.profunctors.uast.structural_test_coverage

MCP<br>topos.mcp.cache

topos.mcp.diagnostics

topos.mcp.evaluation

topos.mcp.formatting

topos.mcp.schemas

topos.mcp.security

topos.mcp.security_findings

topos.mcp.server

topos.mcp.resources.docs

topos.mcp.tools.assess

topos.mcp.tools.compare

topos.mcp.tools.coverage

topos.mcp.tools.docs

topos.mcp.tools.evaluate

topos.mcp.tools.inspect

topos.mcp.tools.preferences

CLI<br>topos.cli.main

topos.cli.diagnostics

topos.cli.evaluation

topos.cli.installation

topos.cli.commands.coverage

topos.cli.commands.quality

topos.cli.commands.system

Configuration<br>topos.config

topos.utils.discovery

topos.utils.tree_sitter

Back to top

View this page

Edit this page

Agentic Cost Savings¶

This case study tests a narrow claim: when an agent starts from structurally<br>cleaner code, does the next feature work take less time, fewer tokens, and less<br>money?

In one controlled experiment, the answer was yes. After a Topos-guided refactor,<br>four follow-on Gemini feature sessions used 32.6% fewer tokens , ran 22.9%<br>faster , and had 24.6% lower estimated feature-session cost than the same<br>features on the unrefactored baseline. Including the upfront refactor, this<br>single run cost more overall.

That is not a universal benchmark. It is one synthetic business-software<br>fixture, and more work is needed before making broad claims.

Why this experiment matters¶

Passing tests are not the end of the story for agent-written code. A feature can<br>work and still make the next feature harder to add.

This case study asks whether that extra cost shows up in the agent loop itself:<br>more reading, more tool calls, more tokens, and more time. Topos is already<br>described elsewhere in these docs; here, the question is narrower. Does a<br>Topos-guided structural cleanup make the next agent sessions cheaper?

The setup¶

The fixture was a small healthcare claims engine. It was synthetic, but it had<br>the texture of business software: payment rules, member eligibility, provider<br>networks, deductibles, coinsurance, audit messages, imports, exports, and tests.

The baseline had:

728 total lines.

10 production and test files.

10 passing tests.

83% line coverage.

One overloaded adjudicator.py file at 264 lines.

Topos identified the intended structural problem. The central adjudicator had a<br>SIMPLE score of 0.0, cyclomatic complexity of 34, and a maximum function<br>complexity of 48. In other words, the code worked, but the main decision point<br>had become the place where future changes would get expensive.

The experiment compared two paths from the same baseline:

Condition

What happened

A: no Topos

Gemini added four features directly to the original code.

B: Topos-guided refactor

Gemini first used Topos to refactor the structure, then added the same<br>four features in fresh sessions.

The...

topos graphs functors evaluation uast probes

Related Articles