Related projects

Projects agentic-job overlaps with, builds on or borrows from, apart from gh-aw, which has a page of its own. Descriptions are from each project's documentation as of September and October 2026; corrections are welcome.

Paperclip

Paperclip (paperclipai/paperclip) is a Node.js server and web UI for running a "company" of agents, with an org chart, goals, budgets and a board of issues. People and agents talk through issues and their comments. An agent wakes on a timer, on assignment or on a mention, and runs through an adapter such as Claude Code, Codex or an HTTP webhook.

A task and message board that drives agents is also what the systems built on agentic-job aim for. Two things differ, and agentic-job is one of them.

The first is where the board is. Paperclip keeps its state in its own PostgreSQL database; external trackers are planned as on-ramps "without becoming the agent control plane themselves". The opposite choice is to make the forge's own issues, pull requests and project board the tracker, so that the state is backed up, audited and permissioned with the code it is about. That is a scheduler's concern and is outside agentic-job, which does not know what a board is.

The second is how agents run, which is agentic-job's part. Paperclip's runtime guide says "Local CLI adapters run unsandboxed on the host machine"; remote sandboxes are optional plugins. Here the sandbox is not optional: every run is a CI job on the forge, as an unprivileged user with no write credentials, whose only effects are safe outputs. Its logs and artifacts are the forge's ordinary run records.

fullsend

fullsend is a framework for autonomous agents working on repositories on GitHub, GitLab and Forgejo, with its own dispatch, sandbox, harness and scripts that apply what an agent produced. cgwalters has argued there for adopting gh-aw's safe-outputs model in place of privileged post-scripts (ADR proposal), and on other forges: "generic forge support is actually going to best done by encouraging agent execution inside a Tekton pipeline or so" (comment). agentic-job's aim that its pieces be usable as a Forgejo Actions job or a Tekton step, and not only as a GitHub Actions workflow, goes the same way.

opendatahub-io/agentic-ci

agentic-ci is a Python CLI that runs Claude Code, OpenCode or Codex non-interactively in CI. It has local, podman and openshell backends, token and cost summaries over OpenTelemetry, and gates before and after a run, such as checks for sensitive files, commit authorship and leaked secrets.

It covers much of the same ground: run one agent in one sandbox and report on it. It drives each agent's CLI directly, where agentic-job uses ACP, and its gates overlap with the safe-outputs check. Offering the ACP harness there as a backend is a possibility nobody has raised with that project.

Agent Client Protocol

The Agent Client Protocol and its adapters are what the harness is built on. They are used as they are, and problems are reported upstream.

NVIDIA OpenShell

OpenShell is a policy-driven sandbox for agents that enforces filesystem, system call and network policy in the kernel and injects credentials only for approved endpoints. It is a candidate for the sandbox and egress policy if it can allow nested rootless podman and /dev/kvm. That has not been tried.

Hive

Hive is a Go program that enumerates GitHub issues and pull requests, classifies them, and dispatches agents in tmux sessions. A deterministic pipeline filters, classifies and gates work before any model sees it, and GitHub access goes through a proxy with scoped tokens. cgwalters raised questions about sandboxing, review of each change and forge-native state in discussion #4955.

Hive is a scheduler and agentic-job is not, so the overlap is with what calls agentic-job. Hive's pipeline describes its filter, classify and gate stages, its queue-depth cadence and token tiers, and compares each idea with agent-board's coordinator (agentic-job#2).

devaipod

devaipod, cgwalters' earlier tool, combined OpenCode, podman, devcontainer specifications and a fine-grained forge MCP server. It was archived in July 2026 in favor of separate unprivileged users and OpenShell. Its related projects page covers OpenHands, SWE-agent, Goose and others, which are not repeated here.

Others

Scion is a multi-agent orchestration platform with containers that are agnostic to the agent, credentials per agent and messaging between agents. Ambient Code is a platform for teams of agents hosted on Kubernetes. Both work at a larger scope than agentic-job, which runs one agent per job and leaves orchestration to whatever calls it.