Digibee, a company that uses Claude Code across its team, has found that coding agents struggle with enterprise integration. While useful for greenfield tasks, they break down when faced with recurring, production-grade integrations that require reliability, auditability, and maintenance.
What coding agents handle well
Coding agents excel under narrow conditions: well-documented APIs, one-off tasks, and low-stakes jobs not in production. A quick script to pull data from a public endpoint or a throwaway ETL job works fine. The problems arise when an integration needs to run repeatedly, be reliable, audited, and maintained by someone other than the original prompter.
Three structural gaps
First, agents start from scratch each time. Pre-built connectors for SAP, Salesforce, or NetSuite embed years of knowledge about sequencing, idempotency, and quirks. An agent re-derives all that from scratch, and suffers from the "lost in the middle" effect: long documentation causes LLMs to drop content from the middle of their context window. Obscure APIs lead to code that quietly fails under load, sometimes months later.
Second, agents produce code, not infrastructure. Integrations need retry logic, failure recovery, credential management, audit trails, monitoring, and alerting. Coding agents produce none of that. Prompting around it piecemeal leaves you maintaining both the integration and hand-rolled infrastructure components. An agent optimized for speed isn't designed to fail safely; a bad write can mean unprocessed payments or orders.
Third, agents don't own what they build. When the person who prompted an integration goes on holiday, the design rationale goes with them. There's no structured artifact—no spec, no mapping document, no record of edge cases. API keys and OAuth tokens need scoping, storage, and rotation; generated code has no opinion on any of it. Scaling to a hundred bespoke integrations yields a hundred codebases to secure and update independently.
The honest assessment
Digibee makes an integration platform, so they have a vested interest, but the technical critique stands on its own. Claude skills can help marginally by embedding documented edge cases, but a skill is a knowledge layer—it can't detect runtime failures, maintain a connector as SAP's idempotency changes, or alert before reconciliation breaks.
The real lesson: productivity gains from coding agents in greenfield development don't automatically transfer to environments requiring operational continuity, governed credentials, and audit trails. Using a coding agent for enterprise integration expecting the same results is a category error.
Recommendations
- For greenfield or one-time integrations: coding agents are genuinely useful. Ship fast.
- For recurring, production integrations: budget separately for retry logic, monitoring, credential management, and audit trails—things the agent won't produce.
- For enterprise-scale integration backlogs: a purpose-built integration platform (AI-native or otherwise) is likely the right tool. The agent speeds up logic; the platform handles everything that keeps it running.
The tools are impressive, but the failure modes are real. Know where one ends and the other begins.