Useful idea. RSS is already structured enough to make this work well, but the real value is not just summarizing. It is filtering what actually matters. A good version should probably answer: What changed? Why does it matter? What should I read fully? That saves more time than summaries alone.

“Better” depends on what you’re using it for. Claude usually feels stronger for long context, writing, and structured thinking. ChatGPT tends to be more versatile across coding, tools, and integrations. In practice, most teams end up using both.

That’s exactly where most programs break. Even when infra is secure by design, the actual risk shows up in everyday workflows. People bypass steps, share data differently, or use tools in ways policies never accounted for. Feels like the gap now is less about defining controls and more about seeing how they behave in real usage.

Good comparison. Feels like MCP vs CLI isn’t just about tooling, it’s about where you want control to sit. MCP gives structure and discoverability, which helps as systems grow. CLI keeps things simple and predictable, which is easier to debug. Most real setups probably end up hybrid. MCP for coordination, CLI for execution.

Good framing. Security assessments only work when they look beyond tooling. A company can have strong products in place and still fail because access reviews are weak, processes are unclear, or people don’t know what “secure behavior” looks like in daily work. The people, process, technology split is still one of the most practical ways to find real gaps.

The dip doesn’t surprise me. As models get tuned for safety, latency, and cost, you often see tradeoffs in code quality and determinism. It’s not just “better model every time,” it’s shifting priorities. For dev workflows, consistency matters more than peak output. A slightly weaker but predictable model is often easier to work with than a stronger but inconsistent one.

Yeah, building them is definitely harder than it looks. Most guides get you to a working setup, but the tricky part starts after that, keeping things consistent once you add multiple steps, tools, and real inputs. That’s usually where things start behaving differently than expected.

The rollback idea is solid. Feels like that’s one of the missing pieces right now, because most agent systems assume forward progress, not safe recovery. Only thing I’ve noticed is rollback works great for state, but gets tricky once actions have external side effects, APIs, data writes, notifications, etc. At that point you can revert the system, but not always the impact. Still, having a clean way to step back is a big shift compared to most setups today.

I’d keep AI close to guidance, not execution. Once it directly touches payments or approvals, the cost of small errors gets too high. Feels safer to let it assist decisions, not make them.

That comparison to shared memory makes sense. What’s interesting is how subtle it can be. One slightly off piece of context early on, and everything downstream still “works,” just not in the way you expect.