Engineering teams make natural early adopters for Claude on your cloud platform: they can integrate an API in an afternoon, they already have the review and testing culture that safe LLM use depends on, and their pain points — review backlogs, stale docs, grinding migrations — are text problems. This article covers the team-level, API-driven uses that an engineering manager can stand up on Amazon Bedrock, Google Vertex AI, Microsoft Foundry, or Claude Platform on AWS; individual coding-assistant products are a separate purchasing decision.
The problem: the work between the code
Most engineering time is not spent writing new code. It goes to reviewing other people's changes, understanding unfamiliar code, updating documentation nobody budgeted for, and mechanical migrations — framework upgrades, API deprecations, dependency bumps — that are important, boring, and endless. These tasks share a shape: large amounts of existing text (code) in, judgment-shaped text out. That is Claude's shape too.
A reference design: Claude comments, CI gates, humans merge
The pattern that keeps this safe mirrors every other function in this series. Claude does advisory language work: a first-pass review comment on pull requests flagging likely bugs, risky patterns, and missing error handling; draft docstrings, changelogs, and architecture summaries; proposed patches for mechanical migration steps. Deterministic tooling remains the source of truth: compilers, linters, type checkers, and test suites gate what ships — the model's opinion never substitutes for a passing build. Humans merge: a model-generated review comment is one more input to the human reviewer, and a model-generated patch is a normal PR that goes through normal review and CI.
A minimal CI review step, using Claude Platform on AWS (which gets new API features same-day):
from anthropic import AnthropicAWS
# Reads AWS_REGION and ANTHROPIC_AWS_WORKSPACE_ID from the environment
client = AnthropicAWS()
msg = client.messages.create(
model="claude-sonnet-5",
max_tokens=2048,
system="Review this diff. Flag probable bugs, security issues, and "
"missing error handling only. If none, say LGTM.",
messages=[{"role": "user", "content": diff_text}],
)
print(msg.content[0].text)The same code runs on Vertex AI or Foundry by swapping the client class, and on Bedrock with the anthropic.claude-sonnet-5 model ID. Sonnet 5 is the everyday choice; escalate gnarly architectural questions to Opus 4.8, and use Haiku 4.5 for high-volume classification like triaging build failures.
Migrations and internal tooling
Migrations are where the 1M-token context window (on Opus and Sonnet) earns its keep: entire modules plus the old and new API documentation fit in one request, so the model can propose consistent changes across many files rather than guessing one file at a time. Treat migration output as a factory for ordinary PRs — small, tested, individually reviewed. Internal tooling is the other quiet win: a script that summarizes the week's incidents, drafts release notes from merged PRs, or answers "which services call this endpoint?" over your service catalog. Tool use — letting Claude call functions you define — is available on all four platforms and is the foundation for these; see Tool Use 101.
Rollout advice and pitfalls
Start with the advisory uses — PR review comments and doc generation — because they are useful even when imperfect and risk nothing. Pin model versions in your tooling and upgrade deliberately with a small eval set of past PRs and known-good reviews, exactly as you would any dependency. Watch the platform feature gaps if your plans include batch jobs or code execution tools: several of those features are unavailable on Bedrock and Vertex AI, available on Claude Platform on AWS, and mostly beta on Foundry — check the feature matrix before committing an architecture to a platform.
Pitfalls: engineers treating review comments as authority rather than input (calibrate the team early — the model is confidently wrong sometimes); secrets and credentials in prompts, which your gateway or pre-commit hygiene should strip; auto-merging anything; and skipping evaluation because "engineers will notice" — they will, inconsistently, which is what evals are for.
Where to go next
Read Tool Use 101: Letting Claude Call Your Systems to go beyond text-in-text-out, and Beyond Python: SDK Options for Your Engineering Stack if your services are not Python. To get the first call working, start at the quickstart.