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

Brandon Waselnuk, co-founder of Unblocked, argues that the primary obstacle blocking reliable autonomous software agents is not intelligence or model capability — it is context. His talk dismantles three widely-held myths and presents hard-won lessons from building a production context engine used by engineering teams at companies with up to 20,000 members.

The talk opens with a striking real-world benchmark: the same coding task was run twice — once with only MCP access to the required SaaS tools, and once with Unblocked's context engine. The naive MCP run took 2.5 hours, consumed 20.9 million tokens, required multiple rounds of human correction, and produced code that compiled but would have broken the production system if merged. The context-engine-assisted run completed in 25 minutes, used 10.8 million tokens, and resulted in a PR a senior engineer approved after a single nitpick — a roughly 6x speed improvement and 48% token reduction.

Waselnuk identifies three myths engineers commonly believe. First, naive RAG over documentation constitutes a context solution. He borrows the concept of 'satisfaction of search' from radiology — the phenomenon where a radiologist stops scanning after finding the first abnormality — to explain why naive RAG agents latch onto the first relevant result and miss critical patterns or existing infrastructure. Second, connecting enough MCPs solves the problem. MCPs provide pipes to data but no understanding or reasoning across it; an agent with MCP access still arrives at a codebase as a 'day one employee' who doesn't know what they don't know, leading to rewritten services that already exist. Third, a 1-million-token context window is the answer. Stuffing a million tokens of undifferentiated content into context leaves the model unable to reason effectively; without entity-relationship structure, the large window just 'sits there.'

The solution Waselnuk proposes is a context engine with six key capabilities: unified system context (reasoning across all systems of record — Slack, GitHub, Jira, Confluence, etc.); targeted retrieval (exhaustive, not satisficing search); conflict resolution (arbitrating between a source file in main and a Slack thread where the CTO says the implementation is wrong — the social graph signals authority); data governance (OAuth-carrying MCP delivery so private Slack DMs are never surfaced to unauthorized agents); personalized relevance via social graphs (knowing which codebases and collaborators a specific engineer touches); and token optimization (compressing the retrieved context into a minimal research packet before handing off to the agent).

Three hard lessons from building the system: (1) access is not understanding — the team wasted time betting that MCP access plus model intelligence would be sufficient; (2) conflicts must be resolved, not passed through — early versions of the engine surfaced conflicting context and let the agent choose, producing unpredictable behavior; (3) caching good answers is dangerous — an answer correct at time T is likely stale by T+24 hours as code and conversations evolve, so caching introduces silent hallucinations.

Waselnuk also demonstrates two open-source artifacts previewed at the conference workshop: a social-graph visualization tool (to be released Monday post-conference) that algorithmically maps engineer contribution nodes, peer review relationships, and area expertise from a Git repo; and a live demo of an Unblocked MCP call where Claude constructed a research query, ran with 'effort: high', received a research packet including factory patterns and library modules, and produced a complete, correct integration plan for a first-class Zendesk integration. His closing principle: 'An agent should write code that feels like it was written by someone who's been on your team for years.'

The talk positions context engines as essential infrastructure for any team beyond roughly 20 engineers attempting headless or background agent workflows, and frames the pattern as complementary to Anthropic's concurrent Claude agent announcements.