Priyanka Phatak led Anthropic’s Claude Managed Agents engineering team at the time of its 2026 workshop. Her work on agent infrastructure addresses a specific engineering problem: how to preserve an agent’s work while replacing the execution environments and model-specific software that carry it out. In her explanation of the team’s architecture, separating session history, the agent loop and code execution made those components independently replaceable.
From core services to agent infrastructure
Earlier in her career, Priyanka worked on Yammer’s Core Services team. Her 2014 account of Yammer engineering describes a product that had outgrown its original Rails application, with smaller Java and Scala services supporting feed delivery, search and document collaboration. She also described cross-functional feature teams that brought engineers together with designers, researchers, analysts and QA, connecting specialized infrastructure work to product development.
Priyanka later held engineering leadership responsibilities at Lyft, including work as a senior engineering manager. At Anthropic, she led the engineering team behind Claude Managed Agents, presenting its architecture alongside colleague Gabe Cemaj in their 2026 workshop.
Separating the agent’s work from its runtime
Priyanka describes an agent as a model running in a loop with tools, supported by a harness and a sandbox. The harness manages the model’s execution; the sandbox provides an environment for running code. Her architectural argument begins with a limitation: harnesses encode assumptions about model behavior, and those assumptions can become stale as models improve. Keeping the surrounding interfaces stable allows the implementation to change.
The team’s first design placed the session, harness and sandbox in one container. That coupling made the container difficult to debug and replace. The redesign separated an append-only session log from the “brain”—the harness and Claude—and the “hands,” the sandboxes and tools. Three choices explain what that separation accomplishes:
Recovery from an external session log: Session history survives independently of the process running the agent loop, allowing a replacement harness to resume work after a failure.
Credentials outside the sandbox: An egress proxy injects credentials into outgoing requests, keeping the secrets out of the environment where generated code executes.
History separate from model context: Preserving the event record outside the context window allows context compaction to change what the model sees without erasing the underlying session history.
These are shared achievements of the team Priyanka led. In her architectural walkthrough, she reported that time to first token fell by about 90% following the redesign. The design also allows one brain to work with multiple execution environments, rather than tying each agent to a single container.
The Claude Managed Agents service exposes this infrastructure through agents, environments, sessions and events. Priyanka introduced those interfaces with a simple joke-teller agent and a console demonstration; Gabe led the subsequent research-desk build. Her defining contribution to the workshop was explaining why the runtime needed to be decomposed: session history must remain useful even as containers fail, context is compacted and model-specific harness decisions change.
Anthropic’s first Managed Agents architecture put the session history, harness and sandbox in one container. Separating them changed failure recovery, credential handling, context management and the path to a first response.
Harness workarounds depend on model behavior. Separating interfaces lets the harness change when those assumptions become stale.
The session log preserves continuity while the harness and sandbox remain replaceable. Sandbox failure returns a decision to Claude; harness failure is recovered by reading the log.
Preserved events support different context strategies, while on-demand sandbox provisioning lets inference begin sooner. Anthropic reports about 90% lower time to first token.
Priyanka Phatak and Gabe Cemaj build from a console-created joke agent to a research desk that reads SEC filings, delegates analysis, saves memory and grades its own deliverables. The Q&A draws the practical boundaries around credentials, orchestration, shared knowledge and private data.
A chat interface sends user.message events to a session and relays the API's SSE stream through its backend to the browser.
The Nvidia task produces both a reusable company memory note and a structured scorecard. Specialists return findings; the main agent creates the deliverables.
Outcomes grade session work against a task-specific rubric and permit retries. They do not provide separate subagent outcomes or the per-session spending caps discussed in Q&A.
Vaults substitute real credentials for placeholders at egress only for allowed hosts. Private connectivity can instead use MCP Tunnels or a self-hosted sandbox with internal network access.
Multiple memory stores can accompany a session. Dreaming consolidates transcripts and memories into updated memory; automatic changes to skills, prompts and agent definitions were outside its demonstrated scope.
Priyanka Phatak and Gabe Cemaj explain how to version agent definitions, choose workspace boundaries, evaluate real behavior and keep a coordinator responsive while custom tools wait.
Keep one agent definition in YAML and synchronize deployed copies through CI/CD, including copies in separate customer workspaces.
Use individual workspaces for the customer data segregation discussed here. Session metadata records customer identity without providing that separation.
A company-research demo connects session fan-out, scheduled work and memory improvement. The Q&A then tests the practical boundaries: evaluation, observability, tenancy, scaling and where the application developer still has work to do.
A custom tool can fan one request into multiple sessions, while each session contains its own multi-agent collaboration and outcome definitions.
A deployment schedules new sessions using an agent definition, initial messages and resources. Memory remains an independent store attached to selected sessions.
Dreaming reviews past session transcripts to improve memory. Its proposed lessons still need task-level evaluations, especially when past experiences disagree.
Workspace membership grants broad access within that workspace. Multi-tenant applications must account for the absence of granular platform RBAC in the version discussed.
Managed Agents handles session infrastructure behind API calls, subject to rate limits. Developers still build the customer product, and on-premises operation was only under consideration.
▶ Watch ↗
