Cloud agent orchestrators in the wild

Summary

The “coordinator + overnight coding fleet” idea is no longer niche: Cursor Projects (Sep 2026 beta) is the clearest productized form — a non-coding coordinator, cloud machines by default, shared project files, and subscriptions to Slack / schedules / PRs. Herdr sits one layer below as an open terminal runtime (agent multiplexer + socket API), not a Projects clone; a growing DIY stack (orchestratr, herdr-factory, herdr-outpost, community skills) turns Herdr into factories and remote dashboards. Vendor clouds (Devin managed teams + Outposts, OpenAI Codex cloud / Agents API, Claude Code Routines, Google Jules, GitHub Copilot cloud agent / Agent HQ) sell sandboxed async work → PR. Open-source “PR factories” that matter for overnight use include OpenHands, LangChain Open SWE, and issue→PR CI patterns (Aider in Actions); Sweep’s hosted service shut down in Apr 2026. Gap for Noa: multi-harness, self-hosted always-on on your boxes, a public/personal fleet status API (personal-site pulse), and Projects-like shared context + subscriptions without vendor lock-in or Privacy Mode / Enterprise plan constraints.

Landscape (notable players)

Player	What it is	Orchestration	Hosting	Strengths	Limits	Links
Cursor Projects	Official long-lived work unit + coordinator	Coordinator plans/delegates; many parallel coding agents; Listening/subscriptions	Cursor Cloud Agents (local agent when needed)	Shared synced context; Slack/schedule/PR; “laptop closed”	Beta; not Enterprise; not Privacy Mode (Legacy); vendor lock-in	blog · docs · Automations · Cloud Agents API
Herdr	Terminal runtime / agent multiplexer	CLI + socket API: panes, `agent start/prompt/wait`, lifecycle states	Self-host (laptop/desktop/VPS); SSH multi-machine; “Herdr Cloud” = connectivity relay (waitlist), not hosted agents	Detach & keep running; multi-harness (Claude/Codex/Cursor/opencode/…); agent-native API	Not a ticket→PR factory by itself; you own orchestration	herdr.dev · github.com/herdrdev/herdr · agent automation docs
orchestratr (`orcr`)	Cross-provider orchestrator on Herdr	Path-addressable agents, wait-on-turn, scopes, cron loops, TS SDK + skill	Self-host; needs Herdr on PATH	Addressable fleet, scheduling, SDK	Early/small project; AGPL + commercial dual license	hkandala/orchestratr
herdr-factory	Autonomous work → merged-PR factory	“Belts” of steps (`work`/`evidence`/`review`/`pr`/`custom`); worktrees; CI watch	Local or remote machine via Herdr; data stays on that machine	Ticket/issue/Sentry/markdown → PR; concurrency caps; ask-human; TUI	Beta / “under construction”; opinionated; `--dangerously-skip-permissions` default for unattended	razajamil/herdr-factory
herdr-outpost	Remote dashboard + relay for Herdr fleets	Polls `herdr agent list`; browser ANSI stream; approve/reject; multi-host SSH	Relay on your box + CF Tunnel/Workers (or Fly/VPS/Tailscale)	Mobile triage, push/Telegram, telemetry	Observability/control plane, not a planner	harlanljones/herdr-outpost
herdr-wrapper	Public empty repo (created 2026-09-26)	Unknown — no README/code yet	n/a	Name matches Noa’s DIY framing	Size 0; do not assume features	dwchoo/herdr-wrapper
Devin (Cognition)	Cloud SWE agent + team of managed Devins	Main session coordinates child Devins in isolated VMs; MCP session API; schedules/playbooks	Cognition cloud; Outposts run tool execution on your VMs/K8s	True multi-agent + isolation; Outposts for data residency	Cost (ACUs); reasoning still in Cognition cloud for Outposts	Manage Devins · Outposts
OpenAI Codex cloud	Async coding in cloud sandboxes	Parallel cloud tasks; GitHub/GitLab/Linear/Slack kickoff; env configs	OpenAI-hosted sandboxes	Parallel background work; review diffs → PR	Per-task, less “months of shared project context” than Cursor Projects	Codex cloud · Agents API
Claude Code Routines	Scheduled / webhook / API cloud Claude Code sessions	One routine = prompt + repo + connectors; GitHub events; `/schedule`	Anthropic web infra	Overnight backlog, PR review, alert triage	Daily routine caps by plan; research preview	blog · docs
Google Jules	Async task pool → plan → PR	Issue label / prompt → Cloud VM → Gemini → PR	Google Cloud VMs	Simple async PR factory; concurrency tiers	Task quotas; less coordinator abstraction	jules.google
GitHub Copilot cloud agent / Agent HQ	Agents run in Actions-powered env on GitHub	Issue assign / Agents panel; custom agents; Claude & Codex in HQ	GitHub-hosted	Native issues/PRs; multi-vendor agents; org controls	Task/PR oriented, not months-long shared Project context	About cloud agent · Agent HQ
OpenHands	OSS SE agent platform (GUI/CLI/SDK/Cloud/Enterprise)	Canvas + Agent Server + Automation Server; sandboxes	Self-host or OpenHands Cloud	Full stack; enterprise options	Ecosystem split across repos; ops cost	openhands.dev · docs
Open SWE (LangChain)	OSS async software factory on Deep Agents + LangGraph	Plan → sandbox implement → PR; review/CI babysit; Slack/Linear/GitHub	Self-host (LangGraph/Docker); pluggable sandboxes (LangSmith, Modal, Daytona, E2B, …)	Closest OSS “Projects-like factory”; threads + sandboxes	Active churn; not taking external issues yet	langchain-ai/open-swe
Sweep	Classic issue → PR bot	Label/title triggers	Hosted shut down ~2026-04-09; OSS self-host docs remain	Historically simple PR factory	Hosted dead; maintenance unclear	sweepai/sweep
Aider + CI	Pair-programmer CLI driven from Actions	Issue-labeled workflows → branch/PR	GitHub Actions runners	Cheap DIY PR loop	Fragile; you own prompts/secrets/gates	e.g. community Actions recipes
MetaGPT / CrewAI / AutoGen	Multi-agent frameworks	Role crews / chats	Mostly local or your infra	Prototyping multi-role flows	Not turnkey overnight PR factories (AutoGen maintenance-mode; MetaGPT more full-cycle sim)	MetaGPT, CrewAI, AutoGen repos
Continue	IDE agent (acquired by Cursor)	—	Local	Historical OSS IDE agent	Not a cloud orchestrator path	continue.dev

Community Herdr catalog (orchestrators, factories, skills): yigitkonur/awesome-herdr — lists e.g. herdr-agent-orchestrator, herdr-orchestrator-skill, herdr-factory-loop-skill, herdr-agent-factory, etc.

Smaller overnight PR-factory experiments also appear (e.g. Nocturne, Sigil, Nightcrawler) — pattern: GitHub Actions or local daemon + worktrees + rate-limited PRs. Treat as examples, not mature products.

Common patterns

Coordinator vs workers
Cursor Projects and Devin “Manage Devins” make the coordinator explicitly non-coding (or clean-slate children) so context doesn’t rot. Herdr DIY often uses one controller agent (or orcr / factory supervisor) that only prompts workers.
Isolation
Worktrees (herdr-factory), VMs (Devin, Jules), Actions runners (Copilot), or pluggable sandboxes (Open SWE, OpenHands). Isolation is how overnight parallelism stays safe.
Shared / durable context
Cursor Projects sync project files across agents/machines. Open SWE binds a thread to a persistent sandbox. Herdr itself persists panes/sessions; “project memory” is still yours to design (files, ~/.orcr/data, factory SQLite, vault notes).
Subscriptions / triggers
Slack, schedules, PR/CI, Linear, webhooks, GitHub labels — present in Cursor Automations/Projects, Claude Routines, Codex integrations, Open SWE, herdr-factory sources (Jira/GitHub/Sentry/markdown).
Human in the loop
Approve diffs/PRs; ask-human cords (herdr-factory); mobile approve on herdr-outpost; draft PRs everywhere.
Cost & concurrency control
Caps on concurrent workspaces/tasks (Jules tiers, factory max_active_workspaces, Devin ACU monitoring, Claude routine daily limits, Copilot premium requests).
Remote presence without laptop
Cloud product VMs or self-host Herdr on a always-on box + SSH / outpost / Tailscale. Herdr Cloud (waitlist) is framed as connecting your machines, not renting agent compute.

Gaps Noa’s DIY idea could fill

Aligned with Coding Todos (Minimal cloud agent orchestrator + personal-site agent status):

Gap	Why it matters
Multi-harness on owned hardware	Cursor/Devin/Jules lock harness + sandbox. Herdr already runs Claude/Codex/opencode/Cursor side by side; wrapper can pick cheapest/best model per task.
Always-on without Cursor Projects constraints	Projects: not Enterprise, not Privacy Mode (Legacy), code in Cursor cloud while running. DIY on a VPS/home lab keeps data residency choices.
Public / personal fleet status API	Personal-site digest already needs overnight agent status — not laptop-local Herdr. Thin status JSON from wrapper/outpost → site “pulse”. Vendors don’t expose a cute public portfolio feed.
*Projects-like shared context you* own**	Sync research artifacts, test recipes, preferences across workers (files/git/object store) without waiting for Cursor’s shared project files.
Subscriptions that match Noa’s tools	Slack + GitHub already in stack; DIY can also watch personal Linear/Sentry/herbarium backlog without paying for every vendor automation SKU.
Cost transparency	Per-agent token/ACU dashboards are weak for individuals (Cursor Admin APIs are Team/Enterprise). Wrapper can log BYOK spend.
Composable layers	Runtime (Herdr) ≠ factory (herdr-factory) ≠ remote UI (outpost) ≠ path orchestrator (orchestratr). A thin “wrapper” that only does coordinator + status + subscriptions may be enough — don’t reimplement belts.

Non-goals (already solved elsewhere): reinventing terminal multiplexing (Herdr); reinventing ticket→PR belts if herdr-factory stabilizes; competing with Devin’s full VM browser QA.

How this maps to Noa’s stack (as documented)

Prior digest: Cursor Projects = product; pstack = quality/orchestrate playbooks — complementary.
Personal-site digest: agent status should come from the cloud orchestrator, not laptop Herdr alone.
Coding Todos MVP sketch: create workspace → start agents → prompt/wait → collect results; decide which Projects-like pieces you own (shared context, subscriptions) vs Herdr.
herdr-wrapper: public repo exists (empty as of research time). Treat as placeholder unless/until code lands; don’t confuse with herdr-factory or orchestratr.

Suggested DIY layering (conceptual, not a build plan):

Subscriptions (Slack/GH/cron) → Coordinator (orcr skill / thin wrapper)
        → Herdr runtime (panes + lifecycle)
        → Workers (Claude/Codex/…) in worktrees
        → Status API → personal site
        → Optional: herdr-outpost (triage UI), herdr-factory (if full PR belts needed)

Key links (bookmark set)

Cursor: Introducing Projects · Projects docs · Automations · Cloud Agents API
Herdr DIY: herdr.dev · orchestratr · herdr-factory · herdr-outpost · awesome-herdr · herdr-wrapper
Vendor clouds: Devin Manage Devins · Devin Outposts · Codex cloud · OpenAI Agents API · Claude Routines · Jules · Copilot cloud agent · Agent HQ
OSS factories: OpenHands · Open SWE

Open questions

Is herdr-wrapper (dwchoo/herdr-wrapper) Noa’s intended repo name, or coincidence? Empty today — confirm ownership/intent before building on it.
MVP: thin coordinator + status API on Herdr, or adopt orchestratr / herdr-factory and only add status/subscriptions?
Where does the always-on box live (home lab vs VPS), and is Herdr Cloud waitlist relevant vs SSH/Tailscale/outpost?
Shared context store: git-synced .project/ files, object store, or vault-backed?
How public is fleet status for the personal site (anonymized counts vs named tasks/PRs)?
Cost policy: hard concurrency caps + model routing — mirror Jules/factory limits or Devin ACU-style budgets?
Does Noa need evidence/review belts (herdr-factory) for overnight merges, or is draft-PR + human review enough?
Relationship to Cursor Projects if/when Noa has access: run official Projects for some work, DIY for multi-harness / privacy / portfolio status?

Research notes / confidence

Primary sources fetched 2026-09-26: Cursor Projects blog+docs, herdr.dev, orchestratr README, herdr-factory README, herdr-outpost README, Devin Manage Devins + Outposts docs, Codex cloud, Claude Routines blog, Jules, GitHub Agent HQ blog, OpenHands docs, Open SWE README, GitHub search for herdr-wrapper/outpost.
Sweep hosted shutdown (~Apr 2026) from secondary reports + Sweep repo still present — treat hosted Sweep as dead.
Continue acquired by Cursor — not a cloud orchestrator path.
Star counts / install numbers on herdr.dev are vendor-claimed; not independently verified.
Did not use Codex tooling for this research (per ask).