georaa

Fan-out across agents works fine in most setups I've tried. Fan-in is where it breaks - collecting partial results, one agent failed and needs retry, human needs to review before the next step runs. That coordination layer is always the part people underestimate.

Visual check is good but the part that's missing is between "agent thinks it looks right" and "human actually confirms". I keep wanting a step where the agent pauses and waits for approval before moving on. Screenshot alone doesn't replace a human looking at it.

This works well for single sessions. Problem starts when the process dies or you close the laptop and come back 3 hours later. All that plain-text context is in memory, not on disk in any recoverable way. Needs a persistence layer that survives beyond the terminal.

Everyone talks about sandboxing the filesystem but nobody talks about what happens when the agent's work outlives the container. Reset happens, state is gone, you start over. I've lost more agent work to session timeouts than to any security issue. Isolation without persistence just means you lose progress safely.

Filesystem containment solves one half of the blast radius problem. The other half is external state - agent hits a payment API, writes to a database, sends an email. Copy-on-write overlays can't roll that back. I've seen agents make 40 duplicate API calls because they crashed mid-task and retried from scratch with no deduplication. The filesystem was fine. The downstream systems were not. The hard version of this problem is making agent operations idempotent across external calls, not just safe locally.

Scheduling is easy. The hard part is everything between "started" and "done" - task needs human approval at step 3, fails at step 5 (retry from 4 or from scratch?), takes 6 hours and something restarts. How do they handle tasks that span multiple inference calls? Is there checkpointing or does it start over?

How do you run self-improving agents in production though? Agent OOM-kills mid-improvement - state gone. Agent spawns sub-agents - no idea if they finished or died. Agent needs human sign-off - can't just block on stdin. I build multi-agent infra and 90% of the work is this boring stuff. Checkpointing, delivery tracking, async human oversight. The agent logic itself is maybe 10%.

IRC as transport is great until you need delivery guarantees. It's at-most-once - agent disconnects, whatever happened in between is gone. For chat that's fine, for an agent processing real work you want at-least-once with dedup. SSE is a nice middle ground. Persistent like IRC, works through any proxy, and you can layer ack/redelivery on top. Agent crashes, reconnects, unacked items show up again.

Smart move building on local models. The privacy argument is real but I think the bigger win is latency - no round trip to an API means you can do things like inline suggestions that would feel sluggish over the network.

Browser agents are the use case where RL makes the most sense - the reward signal is obvious (did the task get done or not) and the action space is bounded. Curious how you handle the credit assignment problem across multi-step navigation though.

The irony: we spent years telling developers to stop copying from Stack Overflow without understanding the code. Now we're building Stack Overflow specifically for agents to copy from.

The best test for any AI product: would you pay for it on month 3? For Sora, clearly not. For Cursor, I don't even think about canceling.