Everything Agents

Guided determinism for agents using AgentScript

Combining LLM flexibility with deterministic gates in Agentforce. Built around one real refund agent.

14 lessons73 min read

The pattern this tutorial teaches has a name in the research literature: it is a neuro-symbolic agent architecture, where a neural decision-maker (the LLM planner) operates inside a symbolic scaffold (typed variables, finite-state topic graph, available when gates). Lesson 12 traces this lineage in depth; see Where this fits in AI research for the full mapping to Kautz's taxonomy, ReAct-style tool use, constrained generation, and the CoALA framing.

Neural choice nested inside a symbolic scaffoldA soft, cloud-shaped neural region (where the language model decides) sits inside a hard rectangular frame whose edges are the symbolic scaffolding: available-when gates, the topic finite state machine, and typed Apex inputs and outputs. The two together form a neuro-symbolic system.TYPED I/OAVAILABLE WHENTOPIC FSMSYMBOLIC SCAFFOLDneural choicethe LLM picks an action

The soft neural choice sits inside hard symbolic gates. More on the research lineage →

Lessons

  1. 1.The scenario6 min

    Read this first. Every snippet, diagram, and bug in the tutorial comes from this one refund agent.

  2. 2.Why this is hard4 min

    Pure-prompt agents stall. Pure-deterministic agents are brittle. Guided determinism is the middle path.

  3. 3.The mental model: four control surfaces5 min

    Topic boundaries, instructions, available-when gates, and after_reasoning. Knowing which to reach for first is the whole skill.

  4. 4.Anatomy of a turn5 min

    What actually happens between user input and assistant reply. Enabled tools are computed before the LLM sees anything.

  5. 5.Invocable Apex: writing actions worth trusting6 min

    Bulk shape, reserved names, truthful nulls, deterministic outputs. Stubs that lie poison the planner.

  6. 6.The toolchain and the dev loop5 min

    Edit, validate, deploy, preview, trace, fix, publish, activate. Don't publish during inner-loop iteration.

  7. 7.Reading session traces6 min

    Traces are the truth. EnabledToolsStep tells you what the LLM saw; VariableUpdateStep tells you why state changed.

  8. 8.Bug 1: the silent permission filter6 min

    The agent runs as the Einstein Agent User, not you. Apex without permset access disappears from enabled_tools with no error.

  9. 9.Bug 2: the planner that lies7 min

    @outputs.X is what the planner says Apex returned. Vague schemas plus sticky context produce convincing fabrications.

  10. 10.Defense in depth5 min

    Schema descriptions, flag-based gates, server nonces, re-fetch on next action. Layer 2-3 anywhere stakes are real.

  11. 11.The free-roam variant: when to relax4 min

    Same Apex and same gates, prompt-only orchestration. Safe because the deterministic surfaces stay intact.

  12. 12.Where this fits in AI research9 min

    Neuro-symbolic Type-2 hybrid. ReAct + constrained generation. Not model-based RL, not Soar.

  13. 13.Glossary3 min

    Every term-of-art used across the tutorial, with one-line definitions.

  14. 14.Sticky-note appendix2 min

    The seven things we wish we had known on day one.

Start with lesson 1The scenario · 6 min