A few weeks ago I started exploring something that made me rethink how we build AI workflows. Most of us naturally reach for Python when building AI systems. I do too. Python gives flexibility, full control, and honestly it powers almost everything in modern AI stacks.But while experimenting, I came across Weft from Weaver Mind AI.Weft is still in development, but the idea behind it caught my attention.

Instead of thinking purely in terms of code execution, Weft focuses more on orchestration — how AI models, APIs, retrieval systems, humans, and tools coordinate together as a workflow.

• Python is amazing at building logic.
• Weft feels like it is trying to improve how AI workflows themselves are structured.

That got me curious.

https://github.com/PreethaRaj/TokenWeaver-Lab/releases/download/v1.0.0/Demo.gif

Experiment Setup

So I decided to experiment with it in one of my projects.

I built a local-first research synthesis demo to compare different orchestration styles:

• Weft-style orchestration
• Traditional Python full-buffer processing
• Staged pipeline executor
  • Retrieval, summarization, and synthesis handled as isolated sequential steps
  • No shared orchestration state
• MapReduce-style orchestration
  • Distributed map stages for retrieval and synthesis
  • Followed by aggregation and reduction steps

The project focuses on:

• Token usage
• Cost visibility
• Orchestration efficiency

Goal

Can changing orchestration patterns alone impact token usage and cost behaviour?

Experiment Design

Some things I experimented with:

• Shared retrieval pipelines to keep comparisons fair
• Token and cost visibility between orchestration approaches
• Deterministic local execution for reproducible experiments
• Research synthesis workflows without introducing API costs into evaluation

Measurement Methodology

Token usage was computed by tracking:

• Total prompt tokens
• Generated completion tokens

across each orchestration path.

Cost estimates were normalized using equivalent model pricing assumptions so orchestration differences could be isolated.

Reduction Formula

((baseline token usage - orchestration token usage) / baseline token usage) × 100

• Baseline: Traditional Python full-buffer execution
• Larger context windows were repeatedly passed between stages

Orchestration approaches that reduced repeated context movement showed measurable token efficiency improvements.

Key Learnings

Lesson 1: Orchestration Matters More Than Expected

We usually optimize prompts or models. But sometimes the bigger opportunity sits in workflow design itself.

In these experiments, orchestration decisions influenced:

• How often context was reloaded
• How retrieval outputs were reused
• Whether intermediate results were shared across stages

Small architectural choices created downstream effects on token consumption and execution efficiency.

Lesson 2: Fair Benchmarking Is Harder Than It Looks

Comparing approaches only works when inputs stay consistent.Even small differences can distort conclusions.

To ensure fairness:

• Retrieval outputs were fixed across runs
• Execution remained deterministic
• All orchestration layers used identical source data

Without this control, token reductions may be misleading.

Lesson 3: AI Engineering Is Becoming a Systems Design Problem

Models still matter.

But so do:

• Pipelines
• Retrieval patterns
• Context flow
• Orchestration decisions

The challenge is shifting from:

“Pick the best model”

To:

Design efficient information movement systems

Key factors influencing performance:

• Context buffering strategies
• Retrieval reuse
• Execution ordering
• State management
• Orchestration topology

Disclaimer

This is not about replacing Python.

• Python remains the foundation
• Weft explores an additional layer: AI workflow orchestration

Acknowledgment

Courtesy: https://github.com/WeaveMindAI/weft

Big thanks to Weaver Mind AI for sharing early ideas in this space.

The project is still evolving — and I’m still learning.

Call for Discussion

Curious if others have experimented with orchestration-first AI development approaches.

Links

GitHub: https://github.com/PreethaRaj/TokenWeaver-Lab
Weavermind: https://weavemind.ai/