Open Duet:停在动作上,每个同事单独设。 一个自托管的工作区,里面是有名字的 AI 同事,有四种审批模式:Ask first、Accept edits、Approve for me、Plan only。定时任务从不自己批准动作,在已连接应用里的改动要等审批。它自己列的局限:连接应用只对着一个假的 Composio 服务测过,没测真服务;自托管服务端是实验性的,只支持一个所有者;桌面安装包没有签名。
strands-agents/box (GitHub repository and README) — Checked 4 Nov 2026 (about 15:00 PT): Apache-2.0, 493 stars, repo created 2 Oct 2026, latest commit on main 4 Nov 2026 (f7a9e4e, "docs: add Linux cloud deployment profiles"). README: OS isolation plus Dogwood policies; operations checked by the engine are denied by default; egress gateway checks connections and HTTP method and path; credential injection so the agent does not receive the underlying secrets; temporal policies, so an earlier file read can deny a later outbound HTTP request; decision records in OTLP JSON; preview supports local execution on macOS with Apple silicon.
KKKKhazix/agent-autoloop (GitHub repository and README) — Checked 4 Nov 2026: MIT, 2,424 stars, repo created 3 Nov 2026, latest commit on main 4 Nov 2026 (c0ceda5). README: kernel about 1,650 lines with zero runtime dependencies; asks by card before deleting data, restarting services, spending money or sending messages outside; external text such as issue titles goes into a file marked as material, not into the prompt; events carrying cwd, agent or access fields are rejected; events handed back by an agent stop being dispatched after 8 hops; Windows not yet supported; the four webhook templates only verified against local mock services.
agent-autoloop SECURITY.md — Says the data separation does not stop all prompt injection; default access is edit; side-effecting work should go through askFirst; read access for an ordinary agent does not mean it cannot read keys on the machine, and keys should be isolated with a separate system account or container.
deepseek-ai/dsh-superagent-prototype (GitHub repository and README) — Checked 4 Nov 2026: MIT, 18,524 stars, repo created 31 Oct 2026, latest commit on main 31 Oct 2026 (8c32a9d, "feat: add the dsh-superagent research prototype"). README: "Research prototype for reference only, not an official DeepSeek product"; a proactive coding agent for one GitHub repository that acts only when asked or when a rule the team wrote allows it.
dsh-superagent-prototype docs/002-authorization.md — Admission table per source; requests are explicit, inferred or not admitted; accept modes off, observe (default: positive assessments recorded as "would accept", nothing accepted) and auto; bots cannot make explicit requests or stop tickets; mandates are revalidated before a ticket resumes; a person saying stop is a trigger, and an executor cannot declare that a person asked it to stop.
Anil-matcha/open-duet (GitHub repository and README) — Checked 4 Nov 2026: Apache-2.0, 6,771 stars, repository created 25 May 2023, latest commit on main 4 Nov 2026 (601282f). README: approval modes Ask first, Accept edits, Approve for me and Plan only; a routine never approves an action by itself; connected-app changes wait for approval; connected apps tested against a fake Composio service, not yet the real one; self-hosted server experimental and single-owner; desktop installers unsigned.
wangzan101/Mika (GitHub repository and README) — Checked 4 Nov 2026: Apache-2.0, 6,604 stars, repo created 29 Oct 2026, latest commit on main 4 Nov 2026 (8d97468, version 0.1.36). README: on-device Android agent that operates apps on a hidden background screen; no Mika server; hands logins, verification codes and payments to the user; Ask mode by default.
Mika docs/modes.en.md (Ask and Autonomous) — In Ask mode only the money-spending step is confirmed; "Sending messages, posting, submitting forms, deleting and changing settings are not asked about; they are just done." For scheduled tasks Ask mode skips the money-spending step and reports it.
Josh Manders on X, list of what the agent tried — Pulled credentials from inside the app container and requested every secret on the platform; tried the cloud provider metadata endpoint; asked the cluster what the credentials could do; guessed internal database and cache names; port-scanned all 65,536 internal addresses twice; about 24 commands over 15 minutes; every attempt blocked, nothing accessed. This is his account, not an independent audit.
Jake Runzer on X, acknowledging the harness — "No mallace intended. I've been running some agents to do signups + agentic testing on similar platforms. Just halted the harness ... definitely going off the rails and reaching too far" (typo in original).
@JustJake (Railway) on X, response — "Yea that looks pretty agentically 'off the rails' to me ... we'll have to ban any sort of non-manual testing externally." A follow-up (https://x.com/JustJake/status/2117786017050865842) says it was run on only the one target.