How Coinbase Orchestrates Context-Aware Agents for Scalable Automation

Enterprises are moving from experimentation to enterprise-wide deployments, but the noise surrounding AI agents and automation often obscures the real wins and the critical lessons leaders should be paying attention to. With this shift comes the new reality of scale exposing the cracks. What once looked like quick wins can now uncover deeper challenges in orchestration, governance, and data discipline. For today’s leaders, the question is no longer whether to deploy AI agents, but how to ensure they deliver sustained business value.

CIO News spoke with Saiesh Prakash, Sr. Manager of AI Automation & Enterprise Integrations at Coinbase, on how to find a way through the noise. With a track record of solving complex integration problems at top-tier technology companies like Nvidia and MuleSoft, Prakash brings a pragmatic philosophy that enterprises don't need AI for AI's sake. Instead, they need a deliberate shift in language that prioritizes end-to-end business results over technical novelty -- what he calls "enterprise automations".

• Moving beyond bolt-on AI: Too many organizations treat AI as a surface-level deployment, expecting it to magically fix underlying process issues. Prakash argued that this approach is destined to fail: "The real value emerges when AI is woven into the fabric of the enterprise through a robust orchestration layer." Enterprises seeking to gain the most from AI should focus their efforts on a more balanced approach that involves equal consideration and management of orchestration tools, structured data, and human oversight.

As to this balance, Prakash shared that certain tasks and decision-making can be done by AI, but not everything can be solved with AI. From his perspective, Prakash sees the greatest advantages of using AI coming from a combination of tools and platforms used in combination to achieve business outcomes. Having humans, agents, tools, and platforms working alongside one another is where orchestration tools such as iPaaS remain essential, serving as the connective tissue between legacy systems and new AI capabilities.

• iPaaS: While AI agents bring intelligence and adaptability, they can't replace the foundational role of integration platforms. For Prakash, iPaaS remains essential for an enterprise AI or automation strategy because it ensures data is orchestrated, filtered, and mapped before AI can add meaningful value on top of it. In other words, AI provides the contextual intelligence to turn that structured data into business outcomes. "This layered model," Prakash argued, "prevents the chaos of siloed deployments and ensures your agents or automation is not just scaleable but accountable."

This blended approach is powered by one non-negotiable element: high-quality, contextual data. The old mantra of "garbage in, garbage out" has never been more relevant as an AI agent's intelligence is not inherent; it is a direct reflection of the information it can access. To move beyond simple tasks, agents must consume a plethora of enterprise-specific knowledge alongside explicit task requirements.

• Data access drives results: Prakash underscored that enterprise search must serve as the backbone of agent intelligence, and we've already seen major vendors like Salesforce's Slack implementing it. In his words, "We need to treat our agents like employees, and likewise, it's very difficult to make the correct decisions without access to the right documents, runbooks, or SOPs." Leaders should be thinking about enterprise search like a RAG model that complements the prompts you're giving an agent.
• Orchestration reduces friction: As enterprises accelerate AI adoption, the challenges shift from proving value to managing scale. As Prakash shared, "Without orchestration, organizations risk creating a fragmented ecosystem where hundreds or even thousands of agents operate in isolation. Orchestration layers reduce this friction by providing visibility and ensuring agents can interact, share data, and be governed within a unified framework."

But as organizations succeed in deploying more agents, they risk creating a new and costly problem. Drawing parallels to the "SaaS sprawl" that created a generation of technical debt and security risks, now with AI, enterprises can easily achieve "agent sprawl" which threatens to unleash thousands of unmanaged bots across the enterprise, leading to redundancy, complexity, and a lack of oversight. The result is a new form of technical debt for enterprises characterized by inefficiencies that immediately impacts overhead at unprecedented speed and scale.

• Automation as architecture: For Prakash, the cure to agent sprawl is clear: enterprises must adopt an approach that treats automation as an enterprise architecture problem, not a series of isolated deployments. In his words, "Left unchecked, scattered bots create redundancy, higher costs, security blind spots, and governance risks that echo the worst lessons of SaaS adoption."

• A unified solution: "It's imperative for enterprise teams to have a specific toolset which can give insights into all the different agents that are out there and also enable us to orchestrate between these agents," Prakash explained, "You need a unified solution rather than deploying agents ad hoc with no usability, reusability, visibility, or accountability."

This integrated strategy ultimately redefines the relationship between humans and machines. Rather than replacing people, AI and automation elevate them. By delegating the heavy lifting of data processing to agents, humans are freed to focus on the high-stakes work that requires their unique judgment.

• The human-in-the-loop: Even with orchestration and data discipline, Prakash stressed that AI is not a closed loop. Human oversight remains a non-negotiable safeguard for enterprises, especially in areas where quality control is required. "If your data is being sent outside the organization or for some business processing, you need to have that quality check by an individual, not totally rely on agents. An agent can do the bulk of the work. They can collect your data, they can optimize your data, and they can orchestrate. But a human becomes important during a final step, especially if you have critical business flows."

Ultimately, leaders must recognize that AI is not a replacement for humans, and the fact that AI is force multiplier only when thoughtfully orchestrated with clean data and human checkpoints. By combining structured data, context-aware agents, and proper governance, enterprises can turn pilots into tangible business results. The enterprise mandate is no longer simply to deploy AI; it's to design systems where humans, orchestration, and context-aware agents operate as one. "All of these systems working as one," Prakash said, "Is where innovation, efficiency, and sustainable value converge."