Agent Building is Easy, Context is Key, Says Unblocked Engineer

Jeff Ng, a founding engineer at Unblocked, delivered a compelling presentation at the AI Engineer World's Fair, asserting that while building AI agents has become remarkably simple, their tendency to confidently err remains a significant challenge. Ng highlighted that what once required a dedicated team and a quarter to develop is now more accessible, thanks to maturing frameworks and cloud infrastructure providers. However, the core issue persists: agents often lack the full context needed to perform tasks accurately.

The Evolution of Agent Development

Ng detailed the complex infrastructure required to build a production-ready agent just six months ago. This included systems for session management, checkpointing and state persistence, orchestration, retry logic, sandbox infrastructure, execution isolation, observability, multi-channel adapters, and replay/debugging. He noted that each of these components often represented a distinct company function or even a separate company in its own right.

Key challenges highlighted from this earlier stage included:

• Checkpoint/State Persistence: Long-lived, stateful agents are vulnerable to data loss from ephemeral infrastructure failures. Without proper state persistence, sessions cannot be resumed, leading to costly restarts and potential duplicated side effects.
• Sandbox Infrastructure: Running agent-generated or third-party code necessitates isolated environments to prevent unauthorized access to secrets or network resources, and to avoid destabilizing shared hosts.
• Observability: Debugging agent failures typically involves piecing together logs and traces from numerous systems, adding significant operational overhead.

Ng emphasized that these infrastructural elements, while necessary, do not inherently improve an agent's capabilities; they are merely the 'taxes' one must pay to deploy an agent.

The Rise of Frameworks and Cloud Primitives

The situation has improved dramatically, Ng explained, with cloud infrastructure players like Cloudflare, Vercel, and AWS providing primitives that, when combined with frameworks such as LangChain, Vercel Eve, and LangGraph, abstract away much of the complexity. This allows developers to concentrate on the core logic of the agent itself.

Defining an agent now boils down to simpler decisions: selecting the model, crafting the system prompt, defining available tools and skills, and specifying the sandbox location. Ng demonstrated this with an example of building an issue enrichment system for Linear, which fetches a ticket, determines its type (feature or bug), searches relevant code, and then formulates a plan. He showcased how such a system, when connected to the right context, could provide actionable recommendations.

The Criticality of Context: The Agent's Blind Spot

However, Ng presented a critical failure case: an agent recommended re-enabling a feature that had previously caused an outage, a fact omitted from the agent's available data. The agent lacked the crucial context from subsequent Slack discussions and a postmortem ticket that explained why the feature was disabled. This, Ng illustrated with an iceberg analogy, represents the 'missing knowledge and intent' that often spans across an organization and its various systems.

He explained that when humans interact with agents locally, they naturally serve as the context layer, providing missing facts and catching errors. But as agents become easier to deploy and operate independently, this human oversight is lost, making context a silent failure point. The intuition and tribal knowledge humans possess need to be captured and provided to agents.

Introducing the Context Engine

This is where Unblocked's concept of a 'context engine' comes into play. A context engine is designed to provide task-relevant information based on the user and the situation, resolving conflicts across data sources, respecting permissions, and delivering synthesized, actionable context rather than raw data. Ng stated, "The gap is not intelligence. It's context."

The context engine works by connecting and modeling various data sources such as documentation, code, tickets, and conversations. From this model, agents receive a refined, relevant, and permission-scoped slice of information.

Context Engine vs. MCP

Ng addressed the question of why a context engine is superior to simply connecting multiple APIs (MCPs). While MCPs provide access to data, they lack the understanding to synthesize and rank it. This can flood agents with irrelevant information, increasing context costs and leaving the agent to resolve conflicting data ad hoc.

By contrast, a context engine synthesizes, ranks, and resolves contradictions, ensuring agents receive decision-grade context. Ng demonstrated this with the Linear ticket example, showing how the context engine provided the agent with the necessary postmortem and Slack conversation details, leading to a correct recommendation to prevent further issues.

Beyond Ticket Enrichment: Broader Use Cases

Ng concluded by outlining broader applications for a context layer, including enhancing code generation quality, improving code reviews by grounding findings in team conventions, providing accurate answers for customer success and sales teams, and identifying root causes in incident management. The core idea is to imbue agents with the institutional and tribal knowledge that is essential for effective operation within an organization.