Context management hits differently when you're running persistent agents. For my setup, the hard part wasn't surfacing the right context. It was maintaining reliable task state between cron runs.
Silent failures where the agent thought a job completed but hadn't. Fixed it with a 4-minute lookback window in cron to catch anything that dropped. Wrote up the broader infrastructure decisions at https://thoughts.jock.pl/p/wizboard-fizzy-ai-agent-interface-pivot-2026 Migrated from a custom task system to an open-source Rails board to reduce surface area. Real-world data access gap you mention is real too. Haven't cracked that one. CAPTCHAs still win.
The execution gap point stuck with me…especially how much of it seems driven by context and tooling rather than intelligence.
In other systems I’ve worked around, more intelligence often just makes things more measurable, not easier to execute — especially when the underlying workflows stay the same.
Curious which constraint you see as primary right now (context, data, or privacy), and whether agents actually fix that or just scale the same friction.
Great question. I first wrote this article a few months ago when ~200k context windows were a real headache. My experience more recently with >= 1M token context windows is that it seems to be enough, for now. Of course, context is just a special form of memory (context is to memory as L1 cache is to L2 cache), and memory is clearly still very broken.
I think it depends on the nature of the task. I've sort of given up on using agents for things that require clean access to realtime data, e.g., shopping or travel planning. I think we'll eventually get there but those industries are relatively slow-moving so it'll take time until there are better, more agent-friendly options.
So for now I'd say memory is probably the biggest obstacle to the things I'm working on.
Context management hits differently when you're running persistent agents. For my setup, the hard part wasn't surfacing the right context. It was maintaining reliable task state between cron runs.
Silent failures where the agent thought a job completed but hadn't. Fixed it with a 4-minute lookback window in cron to catch anything that dropped. Wrote up the broader infrastructure decisions at https://thoughts.jock.pl/p/wizboard-fizzy-ai-agent-interface-pivot-2026 Migrated from a custom task system to an open-source Rails board to reduce surface area. Real-world data access gap you mention is real too. Haven't cracked that one. CAPTCHAs still win.
CAPTCHAs are the downfall of modern digital society 🤮
The execution gap point stuck with me…especially how much of it seems driven by context and tooling rather than intelligence.
In other systems I’ve worked around, more intelligence often just makes things more measurable, not easier to execute — especially when the underlying workflows stay the same.
Curious which constraint you see as primary right now (context, data, or privacy), and whether agents actually fix that or just scale the same friction.
Great question. I first wrote this article a few months ago when ~200k context windows were a real headache. My experience more recently with >= 1M token context windows is that it seems to be enough, for now. Of course, context is just a special form of memory (context is to memory as L1 cache is to L2 cache), and memory is clearly still very broken.
I think it depends on the nature of the task. I've sort of given up on using agents for things that require clean access to realtime data, e.g., shopping or travel planning. I think we'll eventually get there but those industries are relatively slow-moving so it'll take time until there are better, more agent-friendly options.
So for now I'd say memory is probably the biggest obstacle to the things I'm working on.
What's your experience been?