Deep Agents New in Deep Agents v0.6
Use Deep Agents v0.6 to run agents on open‑weight models and reduce costs.
Upgrade to Deep Agents v0.6 and enable harness profiles for open‑weight models.
Summary
Deep Agents v0.6, released on 13 May 2026, focuses on performance, scale, and time‑based improvements. The release introduces a lightweight code interpreter that lets agents write and execute code in memory, eliminating the need for full sandbox overhead. Harness profiles provide per‑model tuning for open‑weight models such as Kimi, Qwen, and DeepSeek, delivering 20×+ lower cost than closed‑frontier APIs. Streaming primitives expose typed, subscribable event projections for messages, tool calls, subagents, and custom events, enabling real‑time UIs. DeltaChannel drastically reduces checkpoint storage by up to 100× for long‑running agents, while ContextHubBackend stores skills, policies, and memories in LangSmith Context Hub for versioned collaboration.
The release also adds support for recursive workflows, programmatic tool calling, and a new harness profile abstraction that can be diffed, versioned, and swapped. These features collectively lower token consumption, cut runtime costs, and improve observability across agent deployments.
Key changes
- Code interpreter for in‑memory code execution and tool orchestration
- Harness profiles for per‑model tuning of Kimi, Qwen, DeepSeek at 20× lower cost
- Typed streaming primitives for messages, tool calls, subagents, and custom events
- DeltaChannel checkpointing that stores only diffs, reducing storage by up to 100×
- ContextHubBackend integration for versioned skills, policies, and memories
- Programmatic tool calling (PTC) that removes round‑trip token overhead
- Recursive workflow support via interpreter‑driven subagent calls