Angela Jiang leads product for Anthropic’s Claude Platform, building the infrastructure developers use to turn models into applications and agents. She previously led API at OpenAI and Connect at Stripe, bringing experience in payments platforms to the challenge of making AI systems useful, controllable, and dependable.
From payments infrastructure to agent platforms
At Stripe, Jiang worked on embedded payments for platforms and marketplaces. Her work included no-code pricing tools that let platforms configure fees and apply them automatically. The mechanism addresses a practical tension in platform design: builders need sophisticated infrastructure, while the businesses using it need to express their own rules without implementing every underlying operation.
Her subsequent leadership of API at OpenAI moved that work into the infrastructure connecting developers with AI models. Now, as product leader for Anthropic’s Claude Platform, she works on the systems surrounding those models: how agents run, access information, coordinate tasks, and check results. Her career across Stripe, OpenAI, and Anthropic connects financial infrastructure with AI infrastructure through a recurring product question: which complexities should a platform handle, and which decisions should remain with its users?
Claude Managed Agents makes that question concrete. It combines a low-level agent harness with managed execution infrastructure, giving developers both the machinery that organizes an agent’s work and the infrastructure on which that work runs. This moves platform responsibility beyond providing a model response to supporting an agent’s execution. Jiang’s platform perspective also emphasizes connection across products: MCP-based connectors and shared standards give agents ways to work with information and capabilities in other systems. Interoperability becomes a mechanism for building an ecosystem in which developers can connect existing tools to agent workflows.
Abstraction that preserves understanding
Jiang’s writing explores what users gain—and what they can lose—when complex capabilities become easier to access. In Composable Markets, she considers reducing the knowledge developers need to use machine-learning capabilities while preserving users’ ability to understand and correct outputs. A useful abstraction must leave people enough understanding and control to recognize a failure and intervene.
Idea Fidelity examines the distance between an initial idea and its implementation. Jiang treats building as a process that develops the idea through concrete choices and tradeoffs. Implementation details help determine what a product means and how it behaves; they are part of the intellectual work of making it.
That concern with control carries into her position on domain-specific verification. Jiang argues that legal and financial applications benefit from customized verification and error-handling logic between model output and execution. A model’s proposed result should pass checks appropriate to the application before the system acts on it, and failures need an explicit handling path. For builders, this makes verification part of the product’s execution design, rather than a quality score considered afterward.
Making computation earn its cost
Jiang and Katelyn Lesse jointly develop a complementary approach in Tokens Should Have Jobs: allocate computation to different responsibilities, then judge the allocation by whether it produces usable work.
Give tokens distinct responsibilities. An executor performs the task; an adviser supplies guidance; a grader checks results against a rubric; and a reflective agent writes lessons into memory for later runs. Their financial-analysis experiments compare these strategies under a fixed token budget, testing whether the organization of computation changes performance beyond the benefit of spending more tokens.
Measure the cost of a correct answer. A profit-and-loss statement with incorrect income or cost figures still requires repair. Jiang and Lesse therefore distinguish partial benchmark accuracy from completely correct outcomes and account for the repeated attempts needed to obtain a usable answer. In their reported experiments, advising and grading improved token efficiency under this analysis. The preferred strategy also depended on how reliably an individual run needed to succeed.
Make feedback specific to the task. Their customer-service example uses a store’s refund rules as a grading rubric. Their recruiting example uses accumulated feedback about candidate fit to improve later attempts. These illustrations show how the roles acquire practical meaning: a grader needs criteria for acceptable work, while reflective memory needs feedback worth carrying forward.
Combine roles through agent orchestration. An executor can consult an adviser, submit work to a grader, retry when necessary, and pass execution records to a reflective agent. Jiang and Lesse describe a direction in which models and platforms become better at constructing these strategies dynamically. The immediate opportunity is to choose and combine responsibilities around the task’s requirements, including its tolerance for errors and the cost of another attempt.
Katelyn Lesse and Angela Jiang show why an agent’s token budget is also an allocation problem: advice, evaluation, and reflection can outperform spending the same allowance entirely on execution.
Treat an agent budget as an allocation across jobs, not merely a quantity of execution tokens.
Control the token allowance when comparing strategies. On the reported benchmark, execution scored 76 and advising scored 89 under the same roughly 600,000-token maximum.
Measure the outcome the user can actually use. For the P&L example, anything short of a perfectly scored answer still requires correction or another run.
Claude Managed Agents supplies the concrete individual-agent layer described in the talk; a meta-harness composes roles above it, while automatic strategy construction remains a longer-term goal.
