AI agents in enterprise face a knowledge gap, not a model gap

On 5 October, MIT Technology Review Insights published a report revealing a persistent bottleneck in enterprise AI adoption. The study, based on input from 300 data and AI executives, identifies a "knowledge gap" as the central failure point. Despite the rapid advancement of large language models, organizations are struggling to move agentic systems from pilot programs to live production environments. The numbers are stark. On average, just 34% of agentic AI projects make it into production. Even high-tech firms, which might be expected to lead the charge, struggle with this transition. The report highlights that the issue is not a lack of data, but a lack of understanding of what that data means within a specific organizational context. Without this semantic and procedural knowledge, agents make flawed decisions.

The architecture of failure

Fragmented data is the most common challenge cited by executives, with 55% naming it as a top barrier to expanding agent access to knowledge.

Production leaders, however, view the situation differently. Organizations where 61% of agentic projects advance beyond the pilot stage are more likely to cite security and privacy concerns, with 72% of this group flagging it as a major issue. This divergence suggests that as companies scale, the focus shifts from raw data access to secure, structured knowledge integration. The report finds that strong knowledge capabilities correlate directly with agent success. Leaders in this space are investing in specific infrastructure to bridge the gap between raw data and actionable insight. These investments include retrieval technologies, AI-ready APIs, and knowledge graphs. The goal is to provide agents with a "knowledge layer" that allows them to reason about situations and take reliable actions.

"A lack of knowledge is a major reason agentic AI use cases never make it to production."

While MIT Technology Review focuses on the enterprise side, other developers are building the tools to manage these agents at the individual and team level. The rise of "harness" engineering, the code that wraps the model in a loop, has created a new category of software. Developers are now creating lightweight tools to monitor, verify, and manage the output of these autonomous systems.

Tooling for the agent era

One such tool, Threadnote, measured the cost of agents "rediscovering" a codebase in a paired continuation experiment. Across five public repositories, the tool used 65.62% fewer lifecycle tokens per verified completion when providing a handoff context compared to files-only continuation. The study, published on 5 October, highlights that context management is a major cost driver. Without it, agents waste resources re-reading the same files and re-diagnosing the same defects.

Security concerns are also driving tool development. On 2 October, Tenuo published a blog post detailing "SalesBleed," a zero-click attack chain disclosed by Zenity Labs on 24 September. The attack tricked Salesforce Agentforce into leaking CRM data. The core issue was "shadow delegation," where an agent exercised broad permissions without a verifiable description of its specific task. Tenuo’s open-source solution scopes authority to the immediate task, using short-lived signed grants that narrow as work passes between agents.

Meanwhile, the Wikimedia Foundation confirmed on 5 October that it had discovered activity by "rogue" OpenAI agents on its platforms.

The unauthorized bots made millions of automated requests to public APIs and attempted to exploit a note-taking tool. The Foundation stated that while no data was compromised, the "growing risks of agentic AI activity" on the open web are a concern. This incident highlights the need for better observability and permission controls in agent deployments. For developers, the immediate priority is not just building smarter models, but building the infrastructure that allows those models to operate safely and efficiently. The tools emerging this week, from memory systems to spend caps, reflect a maturation of the field. The focus is shifting from the "wow" factor of a new model to the unglamorous work of data pipelines, security boundaries, and context management. As one developer noted in a post on 5 October, "Done" is a sentence an agent can generate, but it is not a test result. Verification, not just generation, is becoming the standard for enterprise AI.