Agent success depends on organizational context, not smarter models

Peter Werry from Unblocked argues that agent performance bottlenecks are fundamentally information problems, not model quality issues. Agents behave like new employees with every task, missing years of organizational context like Slack decisions, PR discussions, and architectural rationale. The core claim is that feeding an agent all available context causes distraction and waste,

while a context engine pre-processes what matters. The live demo showed cost differences: a task with Unblocked cost sub-$1 and took one minute, while the same task without context cost more and took two minutes while missing nuances. The compounding effect is key—errors cascade when agents operate without organizational understanding. The demo included

an agent discovering its own context in Slack conversations, which is both impressive and a subtle marketing pitch. The engineering social graph feature maps team expertise and review coverage, surfacing gaps in codebase knowledge. This has genuine operational value for staffing and onboarding. However, the evidence is entirely from Unblocked's own product demo,

and the claimed '50% fewer tokens' is a customer quote without independent verification. Tradeoffs include implementation complexity, potential vendor lock-in for context pipelines, and the risk of agent over-reliance on possibly stale organizational documentation. Teams should evaluate whether building context retrieval in-house is more cost-effective than adopting a dedicated context engine solution.