Context
Harvest builds software for wealth management professionals. Like many groups, it had seen AI initiatives multiply: internal agents, product agents, MCP servers. Two AgentCore initiatives were running in parallel, with different technical choices.
The mission: bring every agentic use case in the group onto a single platform.
The challenge
- Reconcile two incompatible approaches without losing business logic already delivered.
- Define standards every team accepts: framework, repository structure, deployment.
- Secure agents that write into business systems.
- Upskill teams new to AWS and AgentCore.
Our approach
Clear architecture decisions. Harness by default, runtime by exception. A federated multi-repository standard. Python and Strands as the default framework. A shared foundation, agentic-foundation, enriched project after project.
An audit, then a plan. An audit of the 26 existing agents: 24 improvement points identified. A remediation plan in 6 waves over three weeks.
Security by default. Guardrails, scoped IAM roles and observability built into the platform. A Kiro agent, agentcore-builder, to generate agents that meet the standards from the start.
A data-driven lifecycle. Each agent moves toward autonomy at the pace of its results. The e-signature agent reached a 99% success rate and is moving to full autonomy.
Upskilling the teams. An AgentCore training series, video modules, and a talk at the internal Tech Meetup. The platform’s standards become the whole team’s standards.
Results
- A first version of the AgentCore platform in production, presented to the executive committee.
- One simple rule, applied: every team building an agent or an MCP server goes through the platform.
- Shared, documented development standards.
- Agents gaining autonomy based on measurements, not gut feeling.
Key takeaway
Centralizing does not mean slowing down. A shared platform lets teams move faster, because security and observability are already in place.
