Stop babysitting your agents... — Brandon Walsenuk, Unblocked

TL;DR

• The gap is context, not intelligence — Walsenuk borrows Andrej Karpathy’s framing and says today’s agents fail less because they’re dumb than because they start as “a brilliant software engineer” with zero knowledge of your org, code patterns, or systems.

• More MCPs don’t create understanding — Giving an agent access to SaaS tools and code systems still led to a PR that compiled but would have broken production, because the model didn’t know which service already existed or what conventions the team actually uses.

• Naive RAG fails from “satisfaction of search” — He compares agent retrieval to radiologists stopping after spotting one issue on an X-ray: once the model finds a plausible pattern, it often stops searching and misses the real implementation path.

• Bigger context windows don’t solve repo-scale reasoning — Even a 1 million token window isn’t enough if the model can’t reason over entities, relationships, and truth conflicts; Walsenuk says cramming in more text mostly helps “find Waldo” style needle-in-haystack problems.

• A context engine needs conflict resolution, permissions, and personalization — Unblocked’s system reasons across code, docs, Slack, and other systems, weighs who said what (for example, a CTO correcting code in Slack), and only returns data the current user is allowed to see.

• Social graphs are a practical unlock for agent quality — In his demo, an open-source tool maps who reviews whose code, what repos people touch, and where expertise lives, so a vague query can be grounded in the engineer’s actual team, history, and likely code paths.

The Breakdown