@ghumare64: For a while now I've held that the agent harness is a distributed systems problem, and that the chat session was never …

For a while now I've held that the agent harness is a distributed systems problem, and that the chat session was never the right unit of work. DeepSeek's Super-agent writeup is the most complete case for that I've read, and it comes with an ablation. ~300 chat and CI messages per run, ~20 of them real work requests. A single session with a subagent tool and a neutral prompt never delegated once across 5 runs. Every new message interrupted the main loop and p95 hit 29.5 min. Their two-level design, a Router that runs without a Bash tool and only dispatches plus Workers that each own one task, came in at 12.4 / 15.7 min p50/p95. The delegation-first prompt still had one run where the model judged the work "small" and did it itself. When delegation depends on the model's judgment in the moment, you get a long tail. Taking execution tools away from the router fixes it in the architecture. What I'm glad to see measured: 1. Events are the input. IM messages, PR comments, CI results and calendar invites all arrive as Triggers. A human prompt is one trigger among many. 2. Sessions are an implementation detail. People can follow 3 to 5 active sessions at once, so once models handle single tasks well, human attention becomes the bottleneck and the architecture has to absorb it. 3. Fast path in code, slow path in the model. Source rules and allowlists drop noise before any LLM call, and follow-ups on a subscribed thread go straight to the owning Worker. Without that pass-through the Router made ~30% more decisions and used 37% more cache-read tokens. 4. Batching keeps one router viable for longer than you'd expect. From 1 to 40 triggers/min, amortized time per decision fell from 6.4s to 1.6s because queued events share one prefill. 5. Compute, state and execution environment are separate concerns. Stateless loop in the cloud, state in a database, a thin daemon on the machine that holds your uncommitted changes and credentials. 6. The watchdog checks progress as well as liveness. A process can be healthy while the agent is stuck in a loop. Worker and Functions end up as the core nouns. I've spent the last year building on those same primitives, so it's good to see the DeepSeek Harness team land on them too. The cost is in the table: 23.2M cache-read tokens per run against 6.1M for the single session. You buy shorter, steadier completion times with tokens, and for team workloads I'd compromise. The open problem they name is the one I care most about. Chain of Recall over a reference graph covers recent context, but after months of running, long-term memory becomes the main bottleneck. That's the gap I've been talking about memories about, and I expect a lot of next year's harness work to happen there. Anyone building always-on agents will probably converge on something close to this, whatever they call the parts.