Building Reliable Commerce and Coding Agents

Reliable agents need structured context, codebase-wide visibility and guarded runtime tool creation.

Summary

Agent-driven commerce depends on structured product data, hybrid search and enough trust for shoppers to delegate purchases, with agentic commerce projected to reach $1.1 trillion in US retail and $3 to $5 trillion globally. AI coding agents increase output but also accelerate duplicated code, inconsistent standards, brittle dependencies and vulnerabilities across large codebases, making searchable, compiler-accurate code graphs essential. AWS's Strands Agents adds a different capability: agents can create, load and use tools and sub-agents during runtime, including a Hawaii travel planner that generated flight, activities and itinerary sub-agents. The shared requirement is controlled context, with catalog enrichment, codebase visibility, sandboxing, constrained permissions and evaluation preventing more automation from producing weaker results.

The videos

PayPal's experiments show that structured product enrichment and hybrid keyword-semantic search improve agent recommendations, while excessive unstructured content increases hallucinations amid projections of $1.1 trillion in US agentic retail by 2030.

Sourcegraph's Agentic Batch Changes applies and audits changes across hundreds or thousands of repositories, addressing the risks that AI-generated code creates in codebases reaching 50,000 repositories.

AWS Strands Agents lets runtime agents write, load and use tools and sub-agents under sandboxed permissions and evaluation, and its Python version created its own TypeScript version.