Building always On agents
Onyx and Echo walk through Anthropic's guide to always-on agents: what makes them different from ordinary agents, the five primitives behind them, and whether the claim that they're newly practical actually holds up.
Transcript
Onyx Anthropic put out this guide on building always-on agents, and the first thing that made me pause… they say an always-on agent isn't actually always on.
Echo Right, it's the user experience that's always on. The thing itself spends most of its time asleep.
Onyx Wakes up when an event hits, does a run, goes back to sleep. And the whole framing is stateful and proactive. Managing context across sessions, acting without direct supervision.
Echo And the article's central claim is that the category's having a resurgence because models finally got good enough, and cheap enough, to make it practical.
Onyx Cheap enough is doing a lot of work there. Opus 5.5 costs seventy-eight percent less per solved task than Opus 5 on long-horizon coding benchmarks.
Echo Wait, seventy-eight?
Onyx Versus forty percent less on typical workloads. That's the gap.
Echo And that gap is the whole argument. Long-horizon tasks benefit disproportionately from cheaper intelligence, and always-on agents are long-horizon by definition. So the economics crossed a threshold where the surrounding architecture is finally worth building.
Onyx mm.
Onyx The product logic works for me, honestly. If you're already paying for context across sessions, cheaper long-horizon work changes what you can offer a user with no human in the loop.
Echo Sure. Now look at what they actually hand you for the architecture. Five primitives. Event inbox, timers, memory, permissions… and an outbox.
Onyx The outbox barely gets any space.
Echo Listed, never defined. That's a genuine hole, because if work reaches the user through an outbox, that's where half the failure modes live. And they don't tell you how to build it.
Onyx The inbox section is way more concrete, though. Don't let event sources call the agent directly. Write everything to a durable list first, one active worker at a time, deduplicate by source ID.
Echo right.
Echo A message queue for agent runs, basically. Timers split into follow-ups and schedules, with a scheduler polling a separate durable store and firing events into the inbox.
Onyx And memory gets its own store. User instructions, user notes, past messages, task progress. Plus they mention… dreaming, as a way to curate memory across sessions.
Echo Dreaming. Sure. Why not.
Onyx Permissions is where my stomach drops a little. A run holds the user's access and operates unattended. So for each tool call, the agent decides. Proceed, block, or wait for approval.
Echo Hence the credential proxy they recommend, sitting outside the agent's environment. Which mirrors ep 1100, the Trace agents and OpenTelemetry episode. Stack control as the product.
Onyx And the memory stuff matches ep 1167, the GitHub agent skills one. Formalize recurring automations with clear contracts instead of bespoke prompt logic.
Echo Same shape every time. Harness robustness, RL graders, multi-scaffold training, that's the infrastructure carrying the actual weight.
Onyx Here's where I get stuck, though. They say models have better judgment now. More reliable at understanding intent, knowing when to check in. And then they never show an always-on agent deployed in production. Not ONE.
Echo mm.
Echo No customer numbers, no error rates, no latency budgets. It's a best-practices guide from the vendor whose platform benefits if more people build always-on agents on Claude Platform.
Onyx Yeah… vendor material.
Echo And that Opus 5.5 cost figure is vendor-reported, on a benchmark nobody can independently verify. So the central claim, that this is newly practical, rests on economics you take on faith.
Onyx So who actually needs this. If your agent has to act across days or sessions with nobody babysitting it, useful skeleton. If you're just doing request-response chat it's, uh… what's the word…
Echo Over-engineering.
Onyx Over-engineering. Thank you.
Echo I'd push further, even. If you're already managing credentials and sessions for human-driven agents, the always-on layer is mostly new surface area for things to go wrong unattended. The permissions section is the only part that takes that seriously. And even there, auto mode's classifiers blocking dangerous actions? Hand-wavy.
Onyx yeah.
Onyx So it holds up as an architecture sketch. Read it as a proven pattern and it's overgeneralizing.
Echo Test it small, is what I'd do. One inbox, one timer, one memory store, and measure what actually breaks when no one's watching.
Onyx That writeup I'd actually read. Alright, Echo.