Skip to content

langclaw 0.2.0 — workflows and the code interpreter

0.2.0 introduces two major primitives: the @app.workflow decorator for durable multi-step routines, and an opt-in sandboxed code interpreter that lets the agent author and run its own workflows at runtime.

Workflows

A workflow is a typed, multi-step routine the agent can call as a tool, a user can trigger via /workflows run, and a cron job can fire on schedule.

from pydantic import BaseModel
from langclaw import Langclaw

app = Langclaw()

class Brief(BaseModel):
    topic: str
    angles: list[str] = ["overview", "risks", "recent news"]

@app.workflow(
    "research",
    input=Brief,
    max_concurrency=4,
    description="Search several angles in parallel, then synthesize.",
)
async def research(ctx, inp: Brief) -> str:
    ctx.phase("gather")

    findings = await ctx.parallel([
        lambda c, a=a: c.tool("web_search", query=f"{inp.topic} {a}")
        for a in inp.angles
    ])

    ctx.phase("synthesize")
    return "\n\n".join(f"## {a}\n{r}" for a, r in zip(inp.angles, findings))

LANGCLAW__WORKFLOWS__ENABLED=true

Three authoring modes ship in 0.2:

Mode Who writes the body When
python You Typed, testable, deterministic control flow
saved The agent (once) Agent writes a JS body, saved to workflows/<name>.js, loads as a tool
llm_authored (experimental) The model, per run Model writes the body fresh each run from a typed contract

Live progress streams to the channel as the workflow runs — users see phase headers and results as they come in.

Code interpreter (RLM)

An opt-in sandboxed JavaScript eval tool backed by langchain-quickjs. The agent writes a program that can call tools, loop, branch, and fan out — including spawning subagents via tools.task.

uv add "langclaw[interpreter]"
LANGCLAW__INTERPRETER__ENABLED=true

The sandbox is capability-scoped: the agent can only call tools from an allowlist, and RBAC middleware runs before it. No ambient filesystem or network access.

This is what powers saved and llm_authored workflow bodies — the interpreter is the runtime for both.

Matrix channel

uv add "langclaw[matrix]"
LANGCLAW__CHANNELS__MATRIX__ENABLED=true
LANGCLAW__CHANNELS__MATRIX__HOMESERVER_URL=https://matrix.org
LANGCLAW__CHANNELS__MATRIX__ACCESS_TOKEN=...

RBAC: subagents and workflows axes

app.role() now accepts subagents and workflows lists alongside tools. Both default-deny — a role must explicitly opt in:

app.role("analyst", tools=["*"], subagents=["researcher"], workflows=["digest"])
app.role("free",    tools=["web_search"], subagents=[], workflows=[])

What's in 0.2

  • @app.workflow() — durable, typed, multi-step routines
  • ctx.parallel, ctx.tool, ctx.subagent workflow steps
  • Three authoring modes: python (recommended), saved, llm_authored
  • Sandboxed code interpreter (langclaw[interpreter])
  • Matrix channel
  • RBAC subagents and workflows axes on app.role()
  • Named agent display_name support
  • Streaming fix: only model chunks reach the channel (not middleware nodes)