Railway launches Cloud Agent Tasks: machine, conversation and result are separate — reusing a VM doesn't mean memory
krishnan · x · 2026-11-20
Railway's changelog #0316 ships Cloud Agent Tasks, and krishnan breaks down the design in a six-post thread: cloud agents now cleanly separate the machine, the conversation and the saved result — reusing the same agent machine does not mean it remembers the last job.
Key points:
- Hand tasks to cloud agents via MCP, CLI or public API: railway ca task run "Fix the failing test in api/ and open a PR" --repo acme/api returns a task ID; check with task status, follow with --watch, get webhook updates, or use --interactive to let the agent pause for questions.
- Per the docs, naming an existing agent starts a fresh session with no memory; replying to a task or reusing its conversation key continues it — and later tasks on the same key queue behind earlier ones.
- Finished tasks keep their results even after the agent sleeps or is deleted; a new agent spins up a billed VM that sleeps after five idle minutes. The machine is a cost line; memory lives in the conversation.
- The conversation key is therefore a design choice: pair a fix with its regression test on one key, isolate unrelated work; one key for everything stalls unrelated jobs, a new key per message drops needed context.
- The author proposes a pilot: submit a task, disconnect, follow up with its key while running an unrelated task on its own key — do contexts stay apart? "Model forgot" may be an application boundary, not model failure; beware context pollution in long sessions and check against the repo.
- Also: railway ca now prioritizes threads needing your attention, and Agent Connectors reach GA.
More from coding & agent
- Dusk bot: open-source Grok Bot alternative reuses your ChatGPT/Claude/Grok subscriptions, 476 GitHub stars — EXM7777 · 2026-11-20
- Theo's Restaurant Analogy Nails Agentic UX: The Whole Kitchen Waits on Your Fries Decision — seatedro · 2026-11-20
- An Interactive Cat's-Eye World in a Single HTML File — No Images, Models, or Libraries — NathanWilbanks_ · 2026-11-20
- Multi-agent RL recipe turns pass@k into pass@1 on hard chess puzzles — shizhediao · 2026-11-20
- Steering a Local LLM's Attention Directly: Sliders That Shift Activations, Not Prompts — round · 2026-11-20
- Eric Schmidt: founding an agentic AI company is the easiest way to get rich in the next two years — AICopyLab · 2026-11-20