Read summarized version with
About project
Working time:
2025 to 2026
Industry:
Accounting software
The service:
Agentic AI
Overview
The customer is a Canadian accounting software company serving small and mid-sized businesses. Its product team had launched AI agents that categorize transactions, reconcile accounts and answer bookkeeping questions inside the product.
The Challenge
The first agents were built directly on AWS Lambda, which worked well for one agent. By the fourth, every agent carried its own plumbing.
1. Duplicated Plumbing
Each agent had its own session store, credential handling, memory and retry logic.
2. Repeated Security Reviews
Security reviewed the same patterns for every new agent, slowing each launch.
3. Limited Tracing
Debugging a wrong answer meant stitching together logs from several functions.
4. Growing Operating Load
Incidents and cost grew with every new agent while the platform team stayed the same size.
Technology & Approach
We moved the agents to Amazon Bedrock AgentCore one at a time, keeping the agent logic and proving identical behavior before any traffic moved.
Implementation
Phase 1: Inventory and Fit
Inventoried four agents with their tools, data access and hosting, and mapped each to AgentCore Runtime, Gateway, Identity, Memory and Observability.
Phase 2: Migration per Agent
Packaged each agent for AgentCore Runtime, exposed ledger and banking tools through AgentCore Gateway, and moved credentials to AgentCore Identity and conversation state to AgentCore Memory.
Phase 3: Parity Tests
A behavior test set of real requests compared answers, tool calls and latency before and after migration, and cutover happened only when parity held.
Phase 4: Staged Cutover
Traffic moved in stages with a rollback path, followed by a cost and operations review for each agent.
Customization & Client Adaptation
Accounting data and in-product agents set the constraints for every migration step.
One Pattern for Every Agent
New agents now start from the same runtime, identity and tracing setup instead of custom code.
Customer Data Boundaries
Tools act only for the signed-in business, with credentials handled by AgentCore Identity.
End-to-End Tracing
AgentCore Observability traces each request through model calls and tool calls, so a wrong answer can be followed step by step.
Framework Kept Where It Fits
Agent logic moved with minimal change, so the product team kept its development workflow.
If you find this case interesting, we recommend taking a closer look at
The Results
All four agents run on AgentCore with behavior parity proven before each cutover, and new agents now reach production in weeks instead of months because security reviews one pattern instead of one per agent.
Long-Term Value
- A shared agent platform the product team builds on.
- Tracing and cost data for each agent from day one.
- A parity test set reused for future model upgrades.
Business Impact
- Less custom infrastructure to maintain.
- Faster security approval for new agents.
- Fewer incidents caused by plumbing rather than agent logic.
