At QCon AI NY 2025, LinkedIn engineers Prince Valluri and Karthik Ramgopal described how the company is building an internal platform to support AI agents at scale. The talk focused on treating AI agents not as isolated tools, but as participants in a governed execution environment designed for large, long-lived engineering organizations.
The presenters framed the core challenge as one of execution rather than intelligence. LinkedIn’s approach separates human intent from machine execution, using structured specifications instead of free-form prompts. These specifications define what should happen, which tools may be used, how success is evaluated, and which actions are explicitly disallowed, reducing ambiguity that often leads to inconsistent or unsafe agent behavior.
“A spec is how we translate the developer’s intent into something the agent can reliably execute.” – Prince Valluri
A central component of the system is an orchestration layer that manages agent execution end to end. Rather than running agents as single-shot interactions, the orchestrator decomposes work into steps, provisions isolated sandboxes, injects scoped context, invokes tools, and records detailed traces of every action taken. The presenters illustrated this with an example of upgrading a shared library across multiple repositories.
The talk emphasized that orchestration enables determinism and repeatability in otherwise probabilistic systems. By standardizing how tools are invoked and how failures are handled, LinkedIn avoids a proliferation of bespoke, fragile agent implementations across teams. This shared execution model also makes agents observable and interoperable, allowing platform teams to evolve capabilities centrally while product teams benefit from reuse.
Another theme was the distinction between autonomy and authority. Agents are allowed to act freely within bounded environments, such as reading and modifying code. Actions with irreversible consequences, including deployments or merges to main branches, remain gated by humans. According to the presenters, this design preserves human accountability while allowing agents to make meaningful progress without constant supervision.
Human involvement is integrated into the system as explicit checkpoints rather than continuous oversight. Agents can pause execution when approvals or decisions are required, resume once input is provided, and learn from outcomes over time. Feedback from reviews, failures, and prior executions is captured as institutional memory, which is then reintroduced as context for future tasks. The presenters argued that many agent failures stem from missing or stale context, making context engineering a more critical investment than model selection.
“We need a unified open-ended platform so that teams focus on their domain problems and we focus on the system and infrastructure problems” – Karthik Ramgopal
The talk also addressed model choice, cautioning against defaulting to custom-trained models. The presenters suggested first determining whether reasoning is required at all, and if so, whether commercial hosted models can meet quality, scale, and latency needs when paired with retrieval and context augmentation. Custom models and fine-tuning were presented as options of last resort due to their operational overhead.
Several production agents built on this platform were discussed, including background coding agents that translate specifications into pull requests asynchronously, and analytics agents that allow engineers and business users to query data and generate analyses without direct involvement from data scientists.
Developers wanting to learn more can listen to the InfoQ podcast “Platform Engineering: Scaling Agents and MCP at LinkedIn,” which expands on the architectural decisions discussed in the session. They can also watch InfoQ.com for video of the presentation soon.
