Fig. 04 · Multi-agent cost cutting for SMEs

Costrym

Pulls a company's accounting data from Xero, QuickBooks, Zoho Books, sevDesk and Expensify, categorises it, and generates cost-cutting tasks that a person approves before anything is scheduled.

Role
Lead engineer
Team
3 engineers
When
Nov – Dec 2025
Stack
Laravel 12, Inertia + React, LarAgent, OpenAI, Pipedream, Firecrawl, Ably, Stripe
Links
GitHub repo

01The problem

Small companies know they overspend on SaaS, cloud and marketing, but the data sits across accounting tools, and analysing it means a finance person exporting spreadsheets by hand.

02How it works

6 steps

  1. 01 · System moves data

    Ingest

    A Laravel job batch runs one job per provider through Pipedream's managed OAuth and lands everything in one financial-records table.

  2. 02 · AI proposes

    Categorise

    Starts only when the whole batch is done. Twenty records at a time into a fixed list of 18 categories, re-queuing itself until nothing is left.

  3. 03 · AI proposes

    Plan

    An orchestrator proposes 5–7 prioritised tasks with estimated savings.

  4. 04 · Person approves

    Approve

    Every task waits as pending until a person approves it. Approval schedules it by priority band.

  5. 05 · Rules decide

    Route

    Deterministic capability scoring picks the specialist agent for each task: cheap, predictable and easy to debug.

  6. 06 · AI proposes

    Deep analysis

    A five-stage pipeline: baseline, cost decomposition, web benchmarking, cost-efficiency ratio, cost-value alignment. Each stage hands a context document to the next, and the arithmetic runs in tool code, not in the model.

  • AI proposes
  • Rules decide
  • Person approves
  • System moves data

03Hard problems

3 of them
01

Ordering a multi-provider chain

Job batches with a completion callback mean categorisation starts only after every provider finishes, and a self-chaining categoriser stays inside token limits.

02

Routing tasks without another LLM call

Generated tasks arrive without an owner. A scoring function matches task text against each agent's declared capabilities, so routing is free and debuggable.

03

Context across a five-stage pipeline

Each stage writes a structured result into a shared document the next stage reads, and each stage is persisted as its own report so every number can be traced back.

04By the numbers

Tested end to end on synthetic transactions.
agent classes
19
accounting integrations
5
stages in the deep-analysis pipeline
5
from first commit to full pipeline
4 wks

05What I'd do differently

19 agents was more than the problem needed. The baseline, decomposition and ratio stages are pure arithmetic and should be plain code feeding one or two reasoning agents. I'd also lock integrations to read-only actions, with a separate approval for any write.
Next · Fig. 05ARIAAutonomous resort operations