Brains vs Hands: How to Run AI Agents Safely in Production — Viren Baraiya

Highlights

• Production agents function as applications rather than isolated components, executing background tasks, event-driven loops, or long-running coordination

• Separating agent responsibility forces non-deterministic LLMs to act as the brain (planning) while deterministic harnesses act as the hands (execution)

• Agentic harnesses operate as late-bound sagas, dynamically constructing and compiling execution paths at runtime from a fixed set of tools

• Deterministic execution harnesses guarantee side-effect management, idempotency, and mandatory human approval gates for critical actions like cluster restarts

• Conductor serves as an open-source workflow orchestration engine to compile LLM plans into fully runnable, state-tracked workflows

Timeline

1. 00:00 - 03:48

   Expanding Agent Architectures Beyond Basic Chatbots

   Initial agent implementations centered around simple tool-calling loops and customer service chatbots. Scaling agents into production requires background execution, event handling for telemetry and alerts, and multi-agent coordination. Breaking tasks into single-responsibility agents mirrors microservices, preventing hallucinations and creating structured application-level workflows.

   • Production environments require agents to operate as scheduled workers, event-driven listeners, and long-running coordinators
   • Single-agent architectures handling multiple responsibilities suffer higher risks of hallucinations
   • Agent systems reflect microservice patterns by combining single-responsibility units through choreography and orchestration
2. 03:48 - 10:23

   Decoupling Brains from Hands for Deterministic Safety

   Non-deterministic reasoning is useful for analyzing situations and planning steps, but operational actions such as payment processing or infrastructure modification require guaranteed deterministic behavior. The harness tracks state and enforces explicit workflows. When an LLM recommends restarting a cluster, the harness controls the exact sequence, manages idempotency, and routes human approval requests via platforms like Slack.

   • Agent harnesses blend LLM non-determinism with deterministic execution rules
   • LLMs serve strictly as the brain for planning, while the harness serves as the hands for execution
   • Critical system state changes require deterministic safeguards, idempotency, and human-in-the-loop approvals
3. 10:23 - 11:49

   Agentic Harnesses as Late-Bound Sagas

   Traditional sagas require hardcoding every possible state branch upfront. Agentic harnesses act as late-bound sagas where the agent dynamically selects from a deterministic set of tools to construct execution paths at runtime. This approach delivers traditional workflow benefits, including state visibility, failure recovery, and long-running durability, without requiring combinatorial hardcoding.

   • Agentic harnesses implement late-bound sagas by assembling tools into dynamic runtime workflows
   • Runtime workflow compilation eliminates the need to hardcode every possible branching condition manually
   • Durable execution engines preserve visibility and state recovery across long-running agent tasks
4. 11:49 - 16:39

   Demonstration of Self-Planning SRE Agent in Conductor

   A live demonstration showcases an SRE remediation agent analyzing log inputs and generating multi-step plans. A specialized plan-and-compile step converts the LLM's proposed sequence into a structured Conductor workflow. The harness executes the sequence, evaluates system state, and runs a second iteration that intelligently skips redundant actions, such as repeating a data rollback, before verifying recovery.

   • The remediation loop gathers evidence, analyzes logs, and generates structured execution plans
   • Plan-and-compile steps convert LLM text plans into executable Conductor workflows
   • Multi-step evaluations prevent redundant actions across iterative execution loops