As of August 2026, two threads on Hacker News this week are a useful gut check for any business owner being pitched on building an in-house AI agent. The first is a developer who spent roughly a year of side-project time building Tilde, a harness SDK meant to decompose frameworks like OpenClaw and Hermes into reusable cloud API building blocks so people could self-host their own code review agent. It landed at 28 points and 6 comments. The second is an Ask HN thread asking a question that sounds almost embarrassingly basic: why do agent frameworks need a concept of "skills" at all, instead of just an AGENTS.md file pointing to folders of markdown docs? That one pulled 13 points but 12 comments of genuine, unresolved disagreement among engineers.
A year to build the plumbing, not the product
Notice what the Tilde developer actually spent a year on. Not the code review logic itself — that part is comparatively easy to prompt for. The year went into the harness: session management, tool-calling scaffolding, decomposing what other frameworks already solved so it could be re-assembled as "building blocks." That is infrastructure work, and it is the same work every DIY agent project eventually runs into, whether the end goal is code review, customer support, or answering "who's open right now" for a local business.
This is the exact pattern we flagged in our look at what actually runs under the hood of an agent stack: the visible feature is a small fraction of the total build. The harness, the retries, the observability, the auth between tools — that's where the year goes, and it's invisible until you're the one maintaining it.
If engineers can't agree on "skills," don't assume your team will get it right the first time
The second thread matters more than its low point count suggests. Twelve comments arguing over whether "skills" (as used in Claude Code and Codex-style harnesses) are architecturally distinct from a well-organized markdown doc folder is not a sign of a settled, boring technical question — it's a sign the field itself hasn't standardized on core vocabulary yet. When the people building these systems for a living are still debating what a basic primitive is for, that's a strong signal the tooling is immature, the best practices are still shifting, and whatever your developer builds in Q3 2026 may look outdated by Q1 2027.
For a business owner, the risk isn't that your team can't eventually figure it out. It's the meter running while they do — a year of iteration like Tilde's, on a side project with no paying customers depending on it, is a luxury a business with live calls and leads coming in every day doesn't have.
The real question isn't "can we build it," it's "what does the debugging year cost us"
Both threads point at the same fork in the road: build the harness yourself and absorb a year of architectural trial-and-error, or hire a team that already resolved those debates and runs the stack as a service. Neither choice is wrong on its face — it depends on whether agent infrastructure is your product or a tool that should just work while you run your business.
If you're a local business owner evaluating whether to have someone in-house "build an AI agent" for booking, outreach, or visibility, ask a pointed question before signing off on the timeline: how many of the last twelve months of comparable open-source projects were spent on the actual customer-facing feature versus the harness underneath it? Based on what's showing up in these threads, the honest answer is usually "most of it went to the plumbing."
What this means if you're weighing AI marketing or an agent build: a self-built agent stack can absolutely work, but budget for a year of harness debugging before it does — most local businesses need the visibility and booking wins now, not after that debugging cycle. If you want to see where you actually stand before committing either way, start with our free free AI Visibility Report.