Hive

Hive is built for the moment a two-hour job dies at minute forty. Agents park their state to disk and resume from exactly that point, and crash safety is the trait you notice daily rather than the one you demo. Getting there is a clone and a quickstart script that sets up the environments, stores your provider credential and opens a dashboard, so the first colony runs before your coffee cools.

It is young, first released in 2026, and the documentation talks about business processes rather than repositories. Pick it when recovery and parallel work are the problem. Pick LangGraph when you want the flow explicit and hand-drawn.

The risk is packaging. This is a uv workspace, and the README states that running pip install -e . from the root produces a placeholder package after which the runtime does not work correctly. That is the reflex of every Python developer alive, and it fails into a successful import with a broken program. Behind it sits a tracker carrying more than nine hundred open items against a runtime released this year.

Use the quickstart path and only that. Done right: budget enforcement lives in the runtime, with spending limits, throttles and automatic model degradation set per team, agent or workflow, so a runaway gets cheaper instead of louder.

1. There is a single primitive. The Queen is a loop; workers are clones carrying identical tools and the identical model against separate tasks, spawned by calling run_worker while the system is already running, so no graph is compiled ahead of time. 2. Coordination happens through a shared tracker ledger and a persistent plan rather than a buffer handed between nodes. 3. Memory is scoped per agent and evolves through reflexion and learned skills.

The property purchased is uniformity: with exactly one kind of agent, every agent inherits the same recovery and the same observability. No benchmark is published.

Textbook open core. The runtime is given away while the company sells the operated version, positioning itself as the native cloud for agents with a harness, a hypervisor, persistent memory and secure virtual desktops. There is no published number anywhere, only a discovery call, which is what a company offers when it has not decided what it is worth.

The marketing is novel: a site where the community goes long or short on which jobs get automated. Cheap attention, real telemetry. The moat stays thin until hosted customers hold data they cannot move. Likely acquirer: an automation vendor needing an agent story. Position: small, revisit when a price appears.

The demo parks a job and resumes it, which is the right instinct. Two things stall the meeting. Agent context escalates to humans through an account-bound Slack or Telegram channel, so work in progress leaves our perimeter into a messaging product, and that is a data-flow review before anyone touches a pilot. The commercial edition publishes no figure, so sixty of anything is a quote we cannot compare to a competitor.

The free runtime needs Python 3.11 and a workspace per developer, roughly half a day each. Contributors must be assigned an issue before opening a pull request, which says the maintainers are triaging rather than scaling. Not yet.

Apache-2.0, and it is indifferent to which model I point it at. LiteLLM sits under the model layer, so Anthropic, Gemini, DeepSeek, Groq, OpenRouter or the Ollama daemon on my own machine are one setting apart. Credentials go into an encrypted store under ~/.hive/credentials rather than an exported variable I leak into a log. Setup builds separate environments for the core runtime and for tools.

The board lists no client for the tool protocol, which is the single closed-feeling part next to everything else I run. The rest is readable Python with a dashboard I reopen from the command line. I would fork this before waiting on the hosted edition.