Earned Autonomy for AI Agents
The 90% coverage gate is enforced on Hive's own build. The other figures are live across every repo Hive manages — the fleet of public hives below — aggregated over their spokes' contributions. Learn how earned autonomy works.
Already running a hive? Open your dashboard · Learn the concepts
How Hive compares
The ADLC platform you own outright
A new category is forming around the autonomous development lifecycle (ADLC) — agents that triage, code, review, fix, and ship across your repos, continuously. Contenders are everywhere now, from GitHub’s own coding agent to Devin to Microsoft Planner to open-source peers like Fullsend. Where Hive stands apart: it makes autonomy something a repo earns as its test coverage grows (the six-level ACMM ladder), enforces what agents can do at the network layer rather than on the honor system, and gives you a live dashboard over the whole fleet — not just activity buried in GitHub.
| Capability | Hive | Fullsend | GitHub Copilot coding agent |
Devin | MS Planner / Jira Rovo |
|---|---|---|---|---|---|
| Free & open source | ✓Apache‑2.0 · CNCF/KubeStellar | ✓Apache‑2.0 | ✗Paid subscription | ✗Paid / enterprise | ✗M365 / Atlassian license |
| Self-hosted / your infra | ✓K8s, LXC, Compose | ✓Self-host supported | ✗GitHub cloud | HybridPlanning stays in cloud | ✗Vendor cloud |
| Bring your own AI & model | ✓Any OpenAI-compatible; self-host vLLM/llm-d | CLI-basedModel per CLI | ✗Locked to provider | ✗ | Vendor models |
| Multi-agent, specialized roles | ✓Scanner, quality, architect, … | ✓6 agents: triage→code→review→retro | ✗Single agent | SubagentsDelegation, not peers | Assignable agents |
| Enforced permissions at network layer | ✓MITM proxy + scoped tokens | Guarded pathsSandbox, not network-layer | Branch rules | Rule engineAgent-layer | Graph ACL inherit |
| Graduated autonomy levels | ✓ACMM L1–L6, advisory→auto-merge | ✗Human-in-loop, no maturity ladder | ✗ | Allow/Ask/Deny | ✗ |
| Transparent AI cost & budget controls | ✓Per-model cost + weekly token budget | ✗ | ✗Opaque in license | ✗ | ✗ |
| Trajectory / goal-drift review | ✓Intent-vs-transcript, injection-aware | Injection defenseNo drift review | ✗ | ✗ | ✗ |
| Durable work ledger (DAG) | ✓Git-backed beads w/ dependencies | ✗GitHub issues/PRs are the state | ✗ | In-run | Task lists |
| Dashboard & fleet observability | ✓Live dashboard, governor, per-agent metrics, leaderboard | ✗GitHub-native only, no UI | In GitHub | ✓Vendor UI | ✓Vendor UI |
| Human-in-the-loop & audit trail | ✓PR holds, pause/resume, full audit | ✓Human in the loop, transparent | ✓Draft-PR review | ✓ | ✓ |
| Auto plan decomposition + plan review | ✓Live at ACMM L5–L6 · decompose → review gate → stall-replan | Prioritize stageNo review-gated DAG | Spec/plan | ✓Interactive Planning | ✓Work breakdown |
✓ yes · partial · ✗ no. Competitor rows reflect each vendor’s public documentation as of 2026; comparisons are our own.
Hive now closes what used to be the one gap: general-purpose plan decomposition with a review-before-execution gate — live today at ACMM levels 5 and 6. Click Plan on any issue (or add the plan label) and Hive breaks an epic into an ordered sub-task DAG, holds it behind a human plan-review gate, and — when an approved plan stalls — detects the stall and re-kicks the architect to revise it, bounded by a per-plan replan cap. Devin, Planner, and others plan; Hive plans, reviews, and self-corrects — and it’s the option you own outright.
Watch
How Hive works — and why not build your own
A walkthrough of the agents, the governor, and the guardrails that make Hive safe to leave running — and why standing up your own agent loop means re-solving problems Hive already handles.
How it works
Coverage first, then autonomy
Point a hive at your repository and start at L1, where agents only advise. Strong test coverage is what earns the confidence to go further — but the autonomy level never rises on its own. You (the admin) choose to raise it, and only then do agents take on more: filing issues, opening hold-gated PRs, and eventually merging on green. Coverage tells you when it's safe to level up; the decision is always yours. A governor decides minute by minute how hard they work based on your queue depth. Hive doesn’t replace maintainers; it earns its way into the work.
Issue filed
A bug report, feature request, or dependency alert lands on the repo.
Triage
The scanner agent reads, labels, and assigns the issue within minutes.
Fix agent
A specialized agent writes the fix in an isolated worktree, opens a PR.
Coverage gate
The deterministic pipeline gates the merge — build, lint, and a per-package coverage threshold. Roughly 1 in 7 agent PRs is rejected here.
Review
Reviewer agents check post-merge state, GA4 regressions, and invariant drift.
Merge
PRs merge on green CI at the autonomy level you allow. No human approval needed at L6.
Rerun
The loop restarts. Security patches, dep updates, and test fixes happen continuously.
Why Hive
Automate the mundane, reclaim your time
Hive is not about replacing humans. Security patches, dependency updates, test fixes, and issue triage are essential but repetitive. Hive automates them so maintainers can spend time on outreach, adoption, innovation, and strategy — the things that are often neglected.
Get a hive for your project
Hosted on our infrastructure — no Kubernetes, no setup. Request one and invite contributors to run the agents.
Contribute compute to a hive
Lend your AI CLI to run agents for a public project via ClankeR, the contributor relay. Pick a public hive from the table below, run the contribute command, and ClankeR hands your CLI tasks to work — your GitHub username shows up as a live contributor on that hive with the PRs you land, not as a new hive. Earn trust tiers as you complete tasks and climb the leaderboard.
A pipeline, not a prompt
Filtering, classification, and merge-gating are shell scripts that run before any LLM sees the work. Agents only get the judgment calls — and a coverage threshold decides what actually lands.
Get started
Running in minutes
The recommended path starts with a hosted hive — no cluster or infrastructure required. Self-hosting is also fully supported.
Login with GitHub
Authenticate so we can provision your hive and install the GitHub App on your repos.
Install the Hive App
Grant the Hive GitHub App access to the repositories you want maintained.
Choose your autonomy level
Pick an ACMM level from L1 Inception to L6 Fully Autonomous. Start conservative, upgrade anytime.
Connect your AI provider
On each agent card, pick the CLI (Claude Code, Copilot, …) and model, then authenticate — open the agent's terminal and run the CLI's login command, or use the card's Login button. Prefer local? Point agents at your own models via vLLM or llm-d — no provider account needed.
Agents start working
Within minutes, agents begin triaging issues, writing fixes, and opening PRs. Watch from your dashboard.
Safety model
Conservative by default
Every level of autonomy is opt-in. The deterministic pipeline gates every change before any AI or merge touches it.
- CI must pass before any PR merges. No exceptions at any ACMM level.
- Agents cannot push directly to main. All changes go through PRs with full audit trails.
- You control the autonomy level. Start at L1 Inception and upgrade when you’re ready.
- Deterministic pipeline filters before LLMs. Build, lint, test, and policy checks gate every change.
- All agent actions are logged. Full transparency in the dashboard and GitHub activity.
- Hosted hives are isolated. Your code stays in your repos — the hive reads issues and opens PRs.
Bring your own AI
Runs the AI you already have
Each agent picks its own method and model. Use a subscription CLI, an API-key provider, or point agents at your own models — no provider account required for self-hosted inference.
| Method | Type | Models | Authentication |
|---|---|---|---|
| Claude Code | CLI agent | Anthropic (Opus, Sonnet, Haiku) | Anthropic subscription — dashboard Login button or /login in the agent terminal |
| GitHub Copilot | CLI agent | Copilot catalog (Claude, GPT, Gemini) | GitHub device-flow login |
| Gemini CLI | CLI agent | Gemini | Google account or API key |
| Goose | CLI agent | Multi-provider | Provider API key |
| Codex | CLI agent | OpenAI | OpenAI login or API key |
| Aider | CLI agent | Multi-provider | Provider API key |
| Bob | CLI agent | IBM | CLI login |
| vLLM | Self-hosted inference | Your own models, your GPUs | None — no provider account |
| llm-d | Distributed inference on Kubernetes | Your own models, your cluster | None — no provider account |
| LiteLLM new | OpenAI-compatible gateway | Anything behind your LiteLLM endpoint | Endpoint URL + API key |
Active Hives
Want to help? Pick a public hive below and click its Contribute button to run its agents through ClankeR and contribute fixes to a hive of your choice. Your work shows up as tasks completed under your GitHub username, on that hive and on the leaderboard.
Cross-Hive Leaderboard
References
Explore the ecosystem
Reading List
From the KubeStellar Medium — April 2026 onward.