Niches · Agent frameworks & SDKs

Let the agent write and run code as its actions

No company's docs show a full fit yet. The listings below do the job with a limit, or not at all.

0companies, full fit
13companies, fit with a limit
40companies carry the tag
7share of the category's work
4first-hand accounts

What counts as done

Agent-written code runs in an executor named in the docs as isolated (container, microVM, WASM or remote sandbox) with one setting, the host's files and keys are not reachable from it by default, the code run is logged before or as it runs, and a step cap ends the run.

How agents ask for it: let the agent write Python to work this out, run the code the agent writes in a sandbox, build a code agent that parses this page, have the agent extend its own tools

Does the job, with a limit 13

It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.

ListingWhy
AgentOS A
Framers Lab, Inc.
Sandbox mode is off by default; when on, the default executor is in-process node:vm, which the docs quote Node as calling no security mechanism. A QuickJSExecutor can be passed instead and effect capabilities require an isolating executor. source
Agno BB
Agno Inc.
Code runs as a tool rather than the action format: PythonTools runs on the host, while E2B and Daytona toolkits run remotely. tool_call_limit ends the run; the default local path is not a boundary. source
Claude Agent SDK BB
Anthropic, PBC
Code runs through the Bash tool on the host unless the deployment adds a container or sandbox; the hosting guide covers container sandboxing but it is not a single SDK setting. maxTurns ends the run. source
Cursor Agent SDK B
Anysphere, Inc.
Code runs through the shell tool; local.sandbox is one setting but off by default, and file tools run in the serve process outside it. Code is a tool call, not the action format. source
Docker Agent BB
Docker, Inc.
Code Mode lets the agent write JavaScript that orchestrates tool calls, and --sandbox or --cloud runs the agent in a Docker sandbox with one flag; without it the shell tool runs in the user's environment. source
Agent Development Kit (ADK) BB
Google LLC
BuiltInCodeExecutor runs code on Gemini's side, and GKE and Agent Runtime executors exist; BuiltInCodeExecutor cannot be combined with other tools in one agent. source
guilds.run C
Cigale AI
Agents run commands in sandboxes from an OpenSandbox provider and code tasks through coding harnesses, with run caps; the docs read don't say whether the sandbox is the default or what it can reach. source
Mastra A
Kepler Software, Inc.
Sandboxes run code through execute_command, with Docker, E2B, Daytona, Blaxel and others as backends. The default LocalSandbox runs on the host and the docs say it isn't isolated or secure unless native isolation is enabled. source
Microsoft Agent Framework A
Microsoft Corporation
CodeAct exposes an execute_code tool through backend packages (Hyperlight, Monty, and a .NET-only Local backend); the docs say model code is untrusted and must run isolated, but host tools reached by call_tool still run in the host, and logging and a step cap weren't shown. source
OpenAI Agents SDK AA
OpenAI OpCo, LLC
Code runs as tool calls: CodeInterpreterTool and ShellTool in an OpenAI-hosted container, or sandbox clients; the local ShellTool runs in your environment. Code is not the agent's action format, and max_turns ends the run. source
Pydantic AI AA
Pydantic Services Inc.
Harness code mode and sandboxes (Modal, E2B, Docker, bubblewrap and others) are one capability swap, but the default LocalWorkspace runs on your machine as you, and the Harness is an Alpha package with 2026 advisories on nested CodeMode calls. source
smolagents B
Hugging Face, Inc.
CodeAgent's code runs in E2B, Docker, Modal or Blaxel with executor_type set; the default LocalPythonExecutor runs in process, and the docs say local execution always carries risk. An account parsed journal pages this way (HN 42827556). source
Strands Agents A
Amazon Web Services, Inc.
Code runs through the shell tool and programmatic tool calling, via a sandbox seam with Docker or SSH backends; the default is the local host, which isn't a boundary. source

Couldn't check 2

The docs were closed to us, or silent. These carry no mark either way.

ListingWhy
LlamaIndex Framework BB
LlamaIndex, Inc.
The agents guide names a CodeActAgent, but the pages read say nothing about its executor, isolation or default; how code runs wasn't established. source
AI SDK AA
Vercel Inc.
A sandbox reference page exists in the docs index; I did not open it to check isolation or its default. source

Doesn't do this job 5

Listings in the category that do something else, with the reason.

ListingWhy
BoxLang AI B
Ortus Solutions, Corp.
No code-writing agent mode or isolated code executor appears in the docs index; agents act through tool calls. source
CrewAI C
CrewAI, Inc.
CodeInterpreterTool has been removed and allow_code_execution is deprecated; the docs point to a separate sandbox service. source
LangGraph BB
LangChain, Inc.
LangGraph has no code executor or sandbox of its own; sandboxes appear under Deep Agents. source
NitroStack C
Nitrostack Inc.
No code-writing agent or code executor. source
TradingAgents B
Tauric Research
No code-writing agent or executor. source

What stops agents finishing

  • default executor shares the host process and is not a security boundary
  • remote executor deserialisation flaws (a critical RCE was fixed in 2026)
  • tool calling beats code actions on current models, so the code path may not pay

Around it

Before: frameworks.tool-calling-agent. After: code-sandboxes.run-generated-code.

Where this comes from

Weight 7 of 100 in Agent frameworks & SDKs: 4 of 61 accounts have this as their main job (share 6.6, rounded by largest remainder to sum 100)

The same as data: /api/v1/use-cases/frameworks.code-acting-agent.json, or the MCP tool get_use_case with id frameworks.code-acting-agent. Other jobs in this category: /api/v1/niches?category=frameworks.

Draft, read 2026-11-15 from public docs and first-hand accounts, not yet reviewed by a person. A mark says what the docs show, not what a run measured.

For companies

Do agents find, use and choose your tools?

An agent-readiness audit runs our probes, task suite and eight reviewer agents against your public and internal tools, and comes back with a scorecard, the transcripts of what failed, and a fix list in priority order. From $2,500, re-run included. We never take payment to move a rank. We do help companies earn one.