Claude Platform expands beyond a reasoning API
Anthropic said in a July 12 video that Claude Platform has changed significantly over the past six months. In the discussion, Claude Platform engineering lead Katelyn Lesse and product lead Angela Jiang said the platform has moved from a basic API for reasoning and tokens to a fuller agent infrastructure that includes memory, outcomes, dreaming, and multi-agent collaboration.
The two executives framed the update through practical feedback from the Managed Agents product line. Their discussion covered agent identity, communication between agents, and how the harness around agents is changing, along with two customer examples.
Managed Agents feedback centers on memory, outcomes, and dreaming
Jiang said the three ideas that have received the most positive feedback since Managed Agents launched are memory, outcomes, and dreaming. She said developers have focused on the clarity of abstraction, where an agent can keep a clear identity and boundary even as the workflow evolves.
Lesse said agents should eventually have identities independent from users. In the model she described, a user gives the agent an outcome, the agent asks for permissions such as A, B, C, and D, the user can approve A, B, and C while rejecting D, and the agent then creates its own service account to carry out the task. She said that structure would make agent behavior auditable and controllable.
For communication between agents, the two said this can happen through APIs or a lightweight MCP server, “just like human-to-agent interaction.”
Harnesses are getting thinner as meta strategies emerge
Jiang said developer thinking around the harness, the external scaffold around an agent, has shifted over the past few months. Earlier approaches often relied on rigid business-process boxes, where step B could only begin after step A had passed. As models have become smarter and gained deeper tool-calling and reasoning ability, she said harnesses have become thinner and many of the old guardrails built for models can now be removed.
In their place, she pointed to what she called a “meta harness,” a way to combine multiple agent strategies. She said current examples include several agents competing on the same problem, adversarial agents where one proposes and another attacks, and an Advisor strategy where an agent that cannot solve a task directly calls a smarter agent for help.
Lesse added that the future may lean toward hybrid strategies: start with an expansive exploration phase, then narrow down and iterate once a framework has been selected.
Urrea and Shopify River offered as examples
Jiang said the most striking case she had seen came from hackathon winner Urrea. She described a factory setting where one expert had built up 10-20 years of experience and knew machine details deeply. Once that person retires, that knowledge can disappear as well. Urrea uploaded SOPs, machine monitoring signals, and parts manuals into an agent system so the agent could simulate the expert’s judgment. Jiang said the case matters because it creates a backup mechanism for expert knowledge that would otherwise be lost through retirement.
Lesse’s example was Shopify’s internal River system. She said River is not limited to writing code. It covers an end-to-end development workflow that spans PR writing, requirement documents, environment setup, software development, and later QA validation.
She said this suggests engineering teams still look broadly similar, but each engineer is being “turbocharged” by agents. In her description, the structure is moving toward each person orchestrating several Claudes rather than working in a traditional “lead + ticket picking” setup.
What Anthropic says it wants next
Looking ahead, Lesse described an ideal state where a user can tell Claude, “I want this outcome, here is the budget, go,” and the agent handles the rest. She said this is the gap the Claude Platform product line is trying to close, so developers can generate what they want without having to work harder or think through more of the process themselves.