Laura: Harper, have you read the Cursor cloud agent post? Because the core claim is kind of wild — they're saying the single biggest factor in output quality isn't the model, it's the dev environment.
Harper: I skimmed it. The environment-not-model framing is the headline, yeah.
Laura: It's such a product-lens insight. Like, locally you never think about it because your laptop just… is the environment. But in the cloud you have to reconstruct all of that, and when you miss something, you don't get a crash. You get subtle degradation. Output that's just slightly worse, and you blame the model.
Harper: Right. That's the part that's actually interesting to me — the failure mode is quiet. It's not an error you can catch, it's the agent slowly losing the ability to verify its own work because it's missing some dependency or tool you didn't think to install.
Laura: And they say that matters MORE as models get smarter, which is the counterintuitive bit. A year ago models couldn't really use the environment anyway, so who cares. Now they can, so the environment becomes the ceiling.
Harper: Yeah, that tracks. Better models exploit their surroundings more, so the surroundings become the constraint. It's a compute-and-context story — the model will reach for whatever's there, and if the shelf is bare…
Laura: You get a sad agent. They ended up building what they literally call enterprise IT for agents. Secret redaction, network policies, credential management. The whole thing.
Harper: Which is exactly what you'd expect! The moment you give an agent a VM and network access, you're a platform team whether you wanted to be or not.
Laura: Okay, but here's where it gets good from the systems side, and I know you're going to like this — they moved to Temporal for durable execution.
Harper: Oh, I did see that. Yeah, that's a real engineering decision, not a vibes choice.
Laura: They started with work-stealing, which is basically transplanting what works locally to a server. One nine of reliability. Fragile.
Harper: Because work-stealing assumes your workers are stable. The moment a pod dies or an inference provider blips, your agent loop just… vanishes. No recovery, no checkpointing. It's fine for a laptop where you're the only process, it's terrible for a cloud VM that can get preempted.
Laura: So they migrated to Temporal and went past two nines. Now handling over fifty million actions per day across more than seven million unique workflows. And over forty percent of their internal PRs come from cloud agents.
Harper: The Temporal migration is the most honest part of the post, honestly. They basically admit they were about to rebuild all the durable execution primitives — retries, scheduling across machines, durability across node failures — and realized Temporal already does that. Smart, and it's the kind of infrastructure call that looks boring but is actually the whole ballgame.
Laura: That's very much the recipe thing from five seventy-six, right? Like, use the right orchestration primitive for the task shape instead of building one universal loop.
Harper: Exactly. And they also changed the workflow shape — went from eternal agent workflows that just run forever to multiple shorter ones that exit after a single task. Makes version upgrades way easier, because you're not stuck with a long-running workflow on old code.
Laura: The decoupling stuff is also sharp. Agent loop, machine state, conversation state — three separate things they keep independent now.
Harper: Mm-hm.
Laura: The loop lives in Temporal, not on the VM, so they can manage pod lifecycles separately. They can run agents across different pod types — readonly VMs, prewarmed VMs. And on the conversation side, they built an append-only storage mechanism that streams updates to clients, with retry-aware rewind.
Harper: The rewind thing is nice. If a step fails after streaming partial output and then gets retried, the client can detect that, rewind the stream, and show the new data instead of the stale stuff. That's a detail most people would skip and then wonder why their UI shows ghost messages.
Laura: Right? And the getting-out-of-the-way section — early on they didn't trust the agent, so the harness would double-check everything, force a commit, push. As models got smarter, they moved logic out of the harness and into agent-controlled tools.
Harper: The CI Autofix example is good. Old harness: grab job failure logs, write them to the VM itself. New approach: give the agent the GitHub CLI, let it search through the output. The notification to the agent got simpler because you're not hardcoding the workflow anymore.
Laura: And they explicitly say the harness isn't going away — what it CONTAINS is changing. Computer use is the current example of still needing a dedicated subagent with custom prompting and screen recording, but the agent controls when to invoke it.
Harper: That's the right framing. The harness is scaffolding, not a cage. You strip out the deterministic logic as the model earns the trust, but you keep the structure for the things the model can't handle yet.
Laura: The forward-looking bit is what they call self-healing — instead of binary hand-holding versus autonomy, you give the agent tools to understand its own environment and report when something's wrong. Secrets missing, network blocked, whatever. Their autoinstall thing is apparently the first pass at that.
Harper: Which circles right back to your constraint-expressibility take from five sixty-three, honestly. The agent needs to be able to express 'I can't proceed because X' and have that be actionable, not just a silent degradation.
Laura: Yeah, it really is the same pattern — make the failure visible and verifiable instead of letting it compound quietly.
Harper: I will say, though — the post is pretty Cursor-specific. These lessons are real, but they're lessons for teams running agents on VMs at scale with their own orchestration layer. If you're at the stage where you're just wiring up an API call, this is downstream of your problems.
Laura: Totally fair. But if you're the team that's already juggling model pools and agent sessions and wondering why your cloud agents feel worse than your local ones? This is the diagnostic.
Harper: Yeah, no, for that audience this is genuinely useful. The environment-is-the-product framing alone probably saves someone a month of model-blaming.
Laura: Alright, Harper. Temporal's in the show notes if anyone wants to dig into the durable execution side — and the Cursor post itself is worth the nine minutes if you're building in this space.
Harper: Nine minutes of your life you won't get back, but better spent than debugging a work-stealing queue at two AM.
Laura: Fair enough. Talk Thursday?