Recoleta Item Note

Oly – Run AI agents, close your terminal, intervene when it needed from anywhere

Oly is a session-persistent PTY daemon for long-running CLI AI agents: even if you close the terminal, the agent keeps running and notifies you when human intervention is needed. Its value is turning “babysitting an…

cli-agent-runtimehuman-in-the-loopsession-persistenceremote-interventionmulti-agent-supervision

Oly is a session-persistent PTY daemon for long-running CLI AI agents: even if you close the terminal, the agent keeps running and notifies you when human intervention is needed. Its value is turning “babysitting an agent while staring at a terminal” into “asynchronous supervision + intervene on demand,” which is better suited to real software engineering workflows.

  • Long-running CLI agents can get stuck when they encounter y/n, permission confirmations, or uncertain decisions, forcing users to keep the terminal open and stay at their computer.
  • Closing the terminal usually means the session is interrupted, context is lost, or reconnection is required, reducing the practicality of AI agents in real development work.
  • Humans need to retain approval and intervention authority, but should not have to wait synchronously the whole time; this is important for human-AI collaboration and multi-agent supervision workflows.
  • The core mechanism is a background PTY daemon: it takes over and holds agent sessions, so tasks continue running after the local terminal is closed.
  • It buffers and replays output, so when users reattach they can see what happened in the meantime and avoid losing context.
  • It detects states that likely require human input and sends notifications, so users only intervene at critical moments.
  • Users can inject input remotely via commands or a browser (such as sending yes or Enter), without even needing to fully reattach to the terminal.
  • The system supports audit logs, integration with external authentication gateways, and supervisor agents overseeing other agents and escalating decisions to humans, creating a mechanism where “humans are always in the loop without needing to watch continuously.”
  • The text does not provide standard academic benchmarks or quantitative experimental results; there are no datasets, accuracy figures, pass rates, latency comparisons, or ablation numbers.
  • The paper/project claims it allows agent tasks to keep running after the terminal is closed, addressing the typical scenario where a task runs for 20 minutes and gets stuck midway on a y/n prompt.
  • It claims to support remote intervention without interruption: for example, injecting a response directly into a session via oly input <id> --text "yes" --key enter.
  • It claims to support browser access and push notifications, enabling session management and human approval “from anywhere.”
  • It claims to provide complete action auditing and a deployment model with no built-in public network listener, emphasizing controlled exposure through external authentication proxies such as Cloudflare Access and Tailscale.
  • It claims to support multi-node/secondary-node session management and an escalatory workflow where “one agent supervises another agent,” but provides no success-rate, efficiency-gain, or user-study metrics.
Built with Recoleta

Run your own research radar

Turn arXiv, Hacker News, OpenReview, Hugging Face Daily Papers, and RSS into local Markdown, Obsidian notes, Telegram digests, and a public site.