What happened
Brookings researchers Jacob Taylor and Kershlin Krishna published a 30+ page working paper codifying "context-maxxing"—user-controlled, open-source AI deployment architectures that give individuals control over the context (domain knowledge, judgment, intentionality) they bring to generative AI interactions. The paper argues that proprietary AI interfaces (ChatGPT, Claude.ai) are designed to capture user context rather than empower user control, eroding cognitive agency—the capacity for people to think and act with AI in ways that support control, efficacy, and mastery. The authors define five infrastructure building blocks (harness, LLMs via API, context web, security, hosting) and three emerging competencies (specification, orchestration, exploration) that characterize context-maxxing. The work draws on the rapid adoption of OpenClaw, an open-source agent harness that surpassed 100,000 GitHub stars in early 2026 and is now used by millions globally.
Why it matters
This paper reframes AI deployment as a power question, not just a productivity question. For executives and boards weighing build-vs-buy AI strategies, it surfaces a third option—user-controlled infrastructure—and argues that the way AI is deployed determines whether it expands or erodes organizational and individual capability over time. The argument has direct implications for vendor negotiations, internal AI governance, and decisions about where context (proprietary workflows, institutional knowledge, judgment) should reside.
Action needed
Evaluate whether your organization's current AI deployment model gives you meaningful control over context—the knowledge, workflows, and judgment that differentiate your operations. If most AI use runs through vendor-controlled SaaS interfaces, consider piloting user-controlled alternatives (open-source harnesses, self-hosted infrastructure) for high-value knowledge work to test whether they preserve cognitive agency and institutional memory.