New Features in Deep Agents v0.6

Deep Agents v0.6 has released 5 core features centered on the model layer, agent layer, extensibility, and long-running performance.

The Code Interpreter is a lightweight runtime that lets agents compose tools, manage state, and control what gets passed into model context. If the sandbox is a code-first way of working with the environment, the interpreter is a code-first way of working inside the agent loop. This enables Programmatic Tool Calling (PTC), where intermediate results stay in runtime state rather than passing through model context, reducing token consumption and unnecessary model round-trips. This is a pattern Anthropic popularized by adding it as an API behavior, but now it can be implemented on any model, including open-source models, through the interpreter.

Recursive workflows are also possible. An agent can maintain a queue of questions, invoke a sub-agent for the next question, store the result, spawn follow-up tasks from that result, and continue until it has gathered enough evidence to synthesize an answer. This is adjacent to the Recursive Language Models (RLM) concept.

Harness Profiles make open-weight models like Kimi K2.6, GLM 5.1, and DeepSeek V4 viable for production agentic work, at costs 20x+ cheaper than closed frontier models. In internal testing, a harness-layer change alone improved gpt-5.2-codex on Terminal-Bench 2.0 from 52.8%→66.5%, and on tau2-bench, prompts and middleware shifted scores by 10-20 points without any model change. Harness Profiles capture model-specific overrides as named, versionable units so tuning work persists.

Streaming provides a typed event stream via stream_events(..., version="v3") for message text, reasoning blocks, tool calls, state updates, subgraphs, sub-agents, custom channels, and final output. It aligns with the new Agent Streaming Protocol, and the LangGraph SDK exposes remote event streaming via client.threads.stream(...). v1 framework integrations for @langchain/react, @langchain/vue, @langchain/svelte, and @langchain/angular were also released alongside this.

Delta channels optimize checkpoint storage by storing only diffs instead of full snapshots. Depending on conversation length and context size, this can yield a 10-100x storage reduction. In an experiment simulating a 200-turn, multi-file coding session, checkpoint storage accumulated to 5.27GB without delta channels, but dropped to 129MB with delta channels enabled.

ContextHub Backend is a LangSmith-based filesystem that provides versioned storage so that improvements to prompts, skills, and other context persist across runs. Agents can read from and write to Hub repositories, and these writes are recorded as commits with history, review, and environment tagging.