A practical architecture guide for building reliable Agentforce agents with deterministic orchestration, bounded reasoning, explicit state, least-privilege tools, and production-grade engineering workflows.

A risk-based architecture for placing reasoning, grounding, state, actions, and Agent Script controls exactly where failure becomes expensive.

A hands-on engineering guide to deterministic Agentforce workflows with explicit state, action contracts, routing gates, and testable execution paths.

What Y Combinator's open source agent harness gets right, and where it stops short of a Salesforce org.

The failure modes that never show up in a demo, and what they have in common.

How to know an agent is getting better when there is no leaderboard for your org.

Why confirmation gates belong on some actions and nowhere near others.

Why retrieval quality depends more on your schema than your embedding model.

Notes on what actually compounds in agent capability, and what does not.

Choosing the right agent loop for Agentforce workloads.

Picking a runtime for agent tooling when both teams have opinions.

A large context window solves a different problem than the one teams reach for it to solve.

Most agent behaviour is determined before the model reads a single instruction.

Prompts are becoming interfaces. What replaces the prompt as the unit of work.

A short list of problems where a flow, a formula, or a report is the better answer.

Where to put the person, and why the default answer is usually wrong.

What to log so that the question after an incident has an answer.

When splitting into several agents helps, and when it is one agent with extra latency.

The CRM cleanup nobody funded is suddenly the thing blocking the agent rollout.