The sharpest take on agentic AI this week argues we’ve been designing agents wrong from the start. A post making the rounds on Hacker News (175 points and climbing) from the Feldera team makes the case that agents shouldn’t act like chatty colleagues. They should disappear into your software stack like good infrastructure.
The author leans on Mark Weiser’s 1991 essay The Computer for the 21st Century, which argued that the most profound technologies fade into the background. Today’s copilots and chatbots do the opposite. They explain themselves at length, generate walls of text, demand turn-taking dialogue, ask clarifying questions, and need constant supervision. That’s a heavy cognitive tax on the human in the loop.
The Hacker News piece flips the assumption. The fix isn’t smarter prompts or better human-style conversation. It’s software designed to meet agents halfway, embedded inside the systems they operate on rather than bolted on through a zoo of external runners.
The patterns that already work
The post highlights three design patterns that quietly make agents calmer and more effective:
- CLI interfaces: A clean command line lets agent loops interact with systems efficiently and burns fewer tokens than scraping a UI.
- Declarative specs: Configs, schemas, and manifests describe the desired outcome instead of step-by-step instructions.
- Reconciliation loops: Borrowed from Kubernetes. You declare a target state, the system converges toward it, and drift gets detected automatically.
These patterns shift the conversation from “what should I do next?” to “here’s the gap, close it.” Less negotiation, more convergence.
The CDC argument
The most interesting move in the piece is about how data reaches agents. Most systems hand agents tables, dashboards, and CSV exports. The agent then has to poll, run expensive queries, diff results, and guess what changed. That’s slow, expensive, and noisy.
Change data capture (CDC) inverts the model. Instead of agents asking “what’s the state now?”, the database pushes precise updates: inserts, deletes, status changes, each tied to specific records.
The fraud-detection example lands well. In a snapshot model, an agent scans a payments table on a schedule to spot suspicious activity. With CDC, the system streams events as they happen. The agent reacts to “transaction T123 inserted, account A flagged” and acts immediately. No re-evaluation, no diffing, just signal.
Why this matters now
The agentic management space is crowded. AutoGPT, Hermes, and similar wrappers all promise generic agent loops. Most of them sit outside your software, treating your stack as something to be queried and prompted.
What stands out here is the reframing. The argument shifts from “build smarter agents” to “build software that agents can live inside.” That’s a meaningful pivot. The first wave of agentic tooling looked a lot like RPA with better vibes. The second wave will look like infrastructure that exposes the right interfaces, streams the right events, and lets agents stay quiet until something actually needs attention.
For builders, the practical takeaways are clear:
- Audit your software for agent-friendly surfaces. Do you have a real CLI, or just a UI?
- Push toward declarative specs over imperative scripts wherever possible.
- If your system holds data, expose change streams. Polling is a tax.
- Stop measuring agents by how much they explain. Start measuring them by how little supervision they need.
Feldera has skin in the game here, since their query engine ships with these patterns and CDC streams baked in. Take the framing with that context. The underlying argument still holds even if you never touch their product. Calm technology for machines is a better north star than humanlike chatbots.
The full write-up and discussion thread are worth reading at the original source.