Paperclip v2026.626.0: Built-in Hermes, Task Watchdogs, and an Ask Mode

What Paperclip just shipped

Paperclip is an open-source AI agent framework, and its v2026.626.0 release bundles 122 commits from 17 contributors. That's a chunky release, and the headline items are worth breaking down because they touch how you actually run and supervise agents rather than just adding cosmetic features.

The four highlights called out are: Hermes now built in, task watchdogs, an ask mode, and workspace downloads. None of these are marketing abstractions — they map to concrete problems anyone building agent workflows runs into: model access, runaway tasks, human oversight, and getting artifacts out of the system.

A quick note on the version scheme. Paperclip uses date-based versioning (2026.626.0 corresponds to a late-June 2026 build), so you won't find semantic hints about breaking changes in the number itself. Check the release notes and changelog before upgrading if you run this in anything resembling production.

The features that matter for building

Hermes, now built in. According to the release notes, Hermes is now a first-class adapter. You can hire `hermes_local` agents that run on your own machine, or `hermes_gateway` agents that run Hermes remotely through a gateway — with no manual plugin install required. The onboarding for the gateway URL and API key ships with what the notes describe as secure defaults. Practically, this lowers the setup tax: model or tool access that previously meant wiring up an external package is now available out of the box. If you'd already installed external Hermes packages, the notes say those can still override or shadow the built-ins through the Adapter manager, so you're not forced into the bundled version. Task watchdogs. This is the feature that should catch the eye of anyone who has left an agent running unattended. A watchdog is a supervisory mechanism that watches tasks and can act when something goes wrong — stalls, loops, or runs that overshoot. The release text doesn't spell out every knob, so treat the exact behavior as something to confirm in the docs. But the direction is right: agents fail in messy ways, and having a built-in mechanism to catch stuck or misbehaving tasks is more useful than most flashier additions. Ask mode. An ask mode typically means the agent pauses to ask a human before proceeding, rather than charging ahead autonomously. For real automation work, this is the difference between an agent you can trust with consequential actions and one you have to babysit or lock down entirely. The middle ground — agent does the work, checks in at the right moments — is where a lot of practical value lives. Workspace downloads. Getting files and outputs out of an agent's workspace sounds mundane, but it's a real friction point. If an agent generates reports, code, or data, you need a clean way to retrieve it. Better workspace downloads means fewer hacks to extract what the agent produced.

Where it fits, and what to check

Paperclip sits in the agent-framework category — code-first tooling for developers who want to define, run, and supervise autonomous or semi-autonomous agents. That's a different audience than the visual, connector-driven world of Zapier and Make, where the appeal is stitching apps together without writing much code. It's closer in spirit to n8n, which is open source and self-hostable and increasingly popular for teams that want to own their automation stack rather than rent it.

The distinction worth keeping straight: framework versus platform. Paperclip gives you primitives — agents, adapters, watchdogs, an ask mode — that you compose in code. That's more power and more responsibility. You'll likely still bring your own model provider through an adapter; the built-in Hermes gateway is one path to that, but frameworks like this are usually designed to work with whatever underlying models you point them at, whether that's OpenAI, Claude, or something you host yourself. The release notes don't enumerate every supported model, so verify the current adapter list against your needs.

A few honest caveats. This is a fast-moving open-source project shipping large date-versioned releases, which means features can shift between builds and documentation may lag the code. The release page itself is mostly a highlights summary — the specifics of how watchdogs trigger, how ask mode integrates with existing agent loops, and how the gateway handles credentials are things you should confirm before committing. Treat the built-in Hermes gateway's "secure onboarding defaults" as a starting point to review, not a guarantee to trust blindly, especially if you're passing API keys through a remote gateway.

The overall shape of this release is encouraging, though. Instead of chasing bigger claims, it's adding the unglamorous plumbing — supervision, human checkpoints, easier setup, cleaner output retrieval — that separates a demo agent from something you'd actually run against real work. Those are the features that decide whether an agent framework survives contact with production.

If you want help evaluating Paperclip or wiring an agent framework into your own stack, browse the provider directory to find people who can put it to work.