LinkedIn did not build another general coding model. It built CAPT, a company context and workflow layer that gives existing agents approved tools and step-by-step playbooks. A second system, Autopilot for Torch, shows how LinkedIn adds a strict verifier loop when the output must become production machine-learning code.

What is LinkedIn CAPT?

CAPT stands for Contextual Agent Playbooks & Tools. LinkedIn's January 2026 account describes it as a framework that gives compatible agents access to internal code search, data platforms, observability, documents, tickets, and other approved systems through Model Context Protocol.

The company says CAPT powers AI-assisted development for more than 1,000 engineers. It is distributed as a Python package and a local MCP server that editors can connect to automatically. LinkedIn's internal distribution updates it in the background, reducing setup work for individual teams.

CAPT does not replace the agent. It supplies context, tools, authentication, and executable company knowledge that an off-the-shelf agent lacks.

What is a LinkedIn agent playbook?

A playbook is a named, reusable workflow with a purpose, inputs, file references, and ordered instructions. CAPT exposes it as a tool that an agent can discover and call. Playbooks can also call other tools, which lets a team compose a longer job from reviewed pieces.

LinkedIn's experiment-cleanup example is concrete. When an A/B test ends, the workflow must read the result, find every related code path, remove the losing variant and stale flags, and verify that the winner is now the default across services. One playbook combines the experimentation APIs, code search, and cleanup instructions so engineers outside the specialist team can perform the work consistently.

LinkedIn says the pattern also supports creating gRPC services, adding endpoints, debugging crashes, reviewing pull requests, querying data, and investigating production systems.

How does LinkedIn organize company and team knowledge?

Central playbooks cover broadly shared work, such as experimentation cleanup, common debugging, data analysis, review, and observability. Local playbooks live with a repository and describe work specific to that service. CAPT discovers both and presents one surface to the agent.

LinkedIn initially grouped tools into namespaces, but engineers still had to predict which group they would need. It later put a few meta-tools in front of the larger catalog. The agent can search for tools by tag, inspect a tool's input shape, and invoke it without loading every tool definition into the prompt.

The company instruments each tool and playbook call, including repository, success, and workflow identity. Those records help platform owners see which workflows are useful and which need stronger guardrails. The article does not publish an overall task success, acceptance, defect, or time-saved series.

What is LinkedIn Autopilot for Torch?

CAPT makes company knowledge executable. Autopilot for Torch is a separate specialist factory for turning TensorFlow models into production-ready PyTorch models and for other machine-learning engineering work.

Its proof loop is generate, score, hint, and regenerate. The verifier checks whether the model trains, produces stable numbers, preserves source behavior, matches the expected structure, fits LinkedIn's platform, and reaches the relevant quality or performance metric. A failure returns typed, prioritized guidance rather than a vague rejection.

Once a result clears the gates, LinkedIn validates it on development GPU pods and promotes it through Flyte workflows. A tracking console shows conversions, iterations, scores, training jobs, workflow errors, pull requests, and generated artifacts.

LinkedIn reports that generated PyTorch code performed strongly on more than 100 public OpenML tasks, that internal teams use generated code as a starting point for framework migrations, and that auto-tuning improved throughput by more than 10 percent on already optimized LLM workloads. These are first-party results from a specialized ML setting, not proof that CAPT or Autopilot improves all software work.

What remains a human decision?

People choose which knowledge becomes a reviewed playbook, who may call each internal tool, what the verifier rewards, and whether a result should enter production. This last point matters because an agent optimizes the scorecard it receives. A shallow metric can produce a technically passing but operationally weak result.

CAPT and Autopilot are described as internal. LinkedIn links no public source release for either system.

What can a product team copy?

Do not begin by teaching a model the whole company. Encode one high-friction workflow as a small playbook with explicit inputs and approved tools. Put team-specific guidance near the code and shared workflows in a central catalog. Instrument every call. For work with measurable outputs, make correctness a hard gate and turn each failure into specific next-step guidance.