The newest hit on Hacker News is not a SaaS product with a waitlist. It is a repository you clone. Agent Swarm, an open source project from Desplega Labs, resurfaced on Hacker News on October 9, 2026, pitching itself as a self-hosted "company operating system" for AI agents: a system where a lead agent takes in work from Slack, GitHub, email or an API and farms it out to worker agents like Claude Code or Codex, each running in an isolated Docker container.
The timing is no accident. Coding agents have moved from experiments to load-bearing tools over the last two years, and teams running them have discovered the awkward question nobody wants to ask: who is watching the agents? The swarm model answers by adding a management layer, one you can run on your own servers, inspect line by line, and cut off from the internet entirely.
What the swarm actually does
The architecture has four parts. Work arrives from Slack, GitHub or GitLab, email, issue trackers, an API, or the command line. A lead agent plans and delegates. Worker agents do the tasks inside Docker containers, reading from and writing to a persistent shared memory store with vector search and identity files named SOUL and CLAUDE.md that tell each agent who it is supposed to be. Output ships as pull requests or as replies in Slack and email.
The distinguishing choice is harness agnosticism. Rather than tying you to one agent runtime, the lead hands work to whichever harness you configure: Claude Code, Codex, pi, opencode, Cursor, Amp, Grok, Devin, or ACP agents. That makes Agent Swarm a layer above tools like Claude Code rather than a replacement for them, and it is a canny bet in a market where today's favorite coding agent may be next year's legacy tool.
Among the features worth noting: schema-validated task results, so a worker's free-form answer becomes JSON a pipeline can safely consume; deferred tasks that wake when a watched task finishes or a deadline arrives; workflows and schedules for recurring work; and review gates where a human approves before a pull request merges or a reply is sent. A documented set of five orchestration patterns (delegate, fan out, gate on humans, retry, persist state) makes the design choices explicit instead of hiding them inside a hosted service.
The model call is the easy part. The hard parts are deciding when work is done, who reviews it, and where the intermediate state lives.
Why self-hosting is the pitch
Enjoying this story?
Get the five most important stories in tech, every morning. Free.
Agent Swarm is MIT licensed, which is doing real work in its appeal. There is no seat count, no API markup, no feature gate between you and the source. You pay only your own model bills, and your prompts, logs, and agent memory live on hardware you control. For teams that would never paste proprietary code into somebody else's swarm-as-a-service, that is the whole argument.
The project also ships like serious infrastructure. There is a Docker Compose path, an OCI Helm chart for Kubernetes, npm and bun packages, Docker images for the server, worker, and dashboard components, and E2B templates. The HTTP API listens on port 3013 with OpenAPI docs. Per the maintainer's Hacker News post, recent additions include an agent-fs integration (described as a Google Drive rebuilt for agents, also MIT licensed), pages and apps shipped as artifacts, and extensions delivered as inert TypeScript hooks that a trusted lead or operator must activate before they do anything.
The activity numbers are striking, with a caveat. The repository was created on December 19, 2025, and by October 9 had about 874 stars, 116 forks, and more than 2,600 commits, with three releases (v1.165.0, v1.166.0, v1.167.0) landing across October 8 and 9 alone. The caveat: many recent pull requests are authored by a bot account called desplega-bot, which suggests the project dogfoods its own swarm. That is charming as a demo and worth knowing as an evaluator, because commit counts overstate the human review the code has had.
The fine print
Agent Swarm by the numbers
Activity signals from the GitHub repository, as of October 9, 2026.
Note: Many recent PRs are authored by an internal bot account; commit counts overstate human review.
An always-on agent company wired to Slack, email, repositories, and issue trackers has a blast radius. A malicious email or issue could try to manipulate a worker, and persistent shared memory is a place where bad instructions can linger across sessions. The maintainers appear aware of the stakes: recent commits include RBAC fixes that stop workers from attributing tasks to other users and gate stdio MCP server creation at agent scope (both landed October 4), plus telemetry fixes (October 12) that stop test boots from reporting as production installs.
Practical evaluation means scoping tokens to the minimum, keeping production credentials out of workers, checking which actions actually require human approval at the review gates, and setting per-harness budgets, because a busy swarm multiplies token bills. The data retention notes deserve a read too: one integration archives verified inbound deliveries and retains them for 30 days. And with near-daily releases, pin your image tags to a specific version rather than tracking latest, because breaking changes are likely.
What to watch
Agent Swarm sits in a crowded lane. The open questions are the ones every young orchestration project faces: will independent teams publish production reviews, how will the maintainers handle security reports given the number of integrations, will the harness list stay maintained as Claude Code, Codex, and Cursor evolve, and will the cloud offering and the self-hosted one diverge over time?
But there is a larger signal worth naming. The self-hosting world spent the last decade on media servers, password managers, and analytics dashboards. Now the same instincts, own your data, read the source, run it on your metal, are arriving for the most consequential software of the moment: teams of agents that can act. A company-shaped swarm you can clone, audit, and disconnect is either the future of small-team automation or a very ambitious hobby. The MIT license means anyone can find out which, on their own servers.
0 Comments