AWS ships ADOP, an agent framework that leaves production model-free
AWS published a reference architecture on 21 August 2026 that hands data pipeline construction to AI sub-agents, then deliberately keeps those agents out of production. It’s called the Agentic Data Operations Platform, or ADOP, and it targets the weeks that teams spend standing up a single new data source.
The mechanism is specific. ADOP launches a Data Onboarding Agent on Claude Code through Amazon Bedrock, using Claude Code’s Dynamic Workflow feature to spawn specialized sub-agents for each stage of pipeline construction. That comes from the post by AWS authors John Cherian and Nuwan Bandara. Sub-agents cover metadata generation, data ontology deduction, quality checks, ETL transformations and orchestration in Airflow or AWS Step Functions.
Anthropic’s own Claude Code documentation describes a subagent as a specialized assistant that runs in its own context window with a custom system prompt, specific tool access and independent permissions. AWS is using that isolation as a governance boundary rather than a context-saving trick. Each sub-agent gets a narrow job and a narrow set of tools.
The design choice that matters comes next. AWS calls ADOP a build-time accelerator, not a runtime dependency. Agents reason, propose and generate in development, engineers review the output, and CI/CD promotes the generated artifacts into staging and production. Those artifacts are deterministic PySpark, SQL, Airflow DAGs, IAM and Cedar policies, and in the default pattern production runs them without calling a model at all.
General tools make a developer faster. ADOP makes every developer consistent.
John Cherian and Nuwan Bandara, AWS
That’s why the guardrails sit in the architecture instead of the prompt. Sub-agents are constrained by an architectural contract covering tool routing rules, Cedar authorization policies, invariants and inline compliance prompts. Every agent decision is traced through AgentTrace, which records intent, tool selected, outcome and cost, and is publishable to Amazon CloudWatch or an OpenTelemetry sink. A Decision Engine encodes your enterprise architect’s standards so different builders don’t produce different architectures.
The rollout plan is where AWS gets concrete, and it isn’t measured in hours.
| Phase | Weeks | Scope |
|---|---|---|
| Phase 1 | 1 to 3 | Two to three engineering champions, one non-critical data source |
| Phase 2 | 4 to 6 | Full platform team, 3 to 5 additional sources of increasing complexity |
| Phase 3 | 7 to 12 | Organization-wide rollout, new sources flow through ADOP |
The catch is that the speed claim carries no number. The post says only that teams running this pattern “have seen source onboarding timelines compress significantly on subsequent sources”, and it labels that directional. There’s no named customer, no baseline and no percentage. Read the title as a design goal rather than a measured result.
AWS is unusually blunt about the failure mode. The post warns that large language models can produce plausible but incorrect logic, and that generated masking rules, retention policies or access controls might be incomplete or subtly wrong. It tells customers to treat every generated control as unverified until legal or compliance validates it. Agent output is a draft, not a certified implementation.
That review burden is the open problem, and researchers are circling it. A paper posted to arXiv on 18 August, AdaLens, argues that conventional interfaces no longer support two requirements of long-running agentic analysis: observability into an agent’s evolving reasoning, and steerability to redirect low-value work mid-run. Its authors evaluate a storyline-based interface through two case studies and a user study. The same gap shows up in how long-horizon agents fail.
Governance is also the reason the multi-tool framing appears at all. ADOP is written so Claude Code, Kiro, Cursor and Codex operate from the same architectural contract, which follows the same logic as Meta’s bet on parallel sub-agents. The reference implementation targets AWS, though the framework extends to services with a CLI or Model Context Protocol interface.
The code sits in a public aws-samples repository. Watch for the number AWS didn’t publish: a first-pass artifact acceptance rate from a named customer would tell you whether reviewing agent output costs less than writing the pipeline yourself.
Get the daily rundown
One email each weekday with the AI news that matters, every claim linked to its primary source.
Free, one email each weekday, unsubscribe in one click. We never sell or share your address.
