Agent rules
A free-text system prompt is one blob where identity, prohibitions, tone and edge cases all blur together. You can't version a sentence, toggle a constraint, or audit what changed — and as the prompt grows, contradictions creep in that silently degrade the agent.
Rules replace that blob with typed sections that compile deterministically into the prompt every LLM node inherits.
Author them under Dashboard → Agents → <agent> → Rules.
Sections
Rules are compiled in a fixed order, so the output is stable:
| # | Section | What belongs here |
|---|---|---|
| 1 | IDENTITY | Who the agent is |
| 2 | MISSION | What it must achieve |
| 3 | PROHIBITION | What it must never do |
| 4 | OBLIGATION | What it must always do |
| 5 | CONDITIONAL | If–then behaviour |
| 6 | TOOLING | How to use its tools |
| 7 | EDGE_CASES | Unusual situations and how to handle them |
| 8 | TONE | How it should sound |
| 9 | PREFERENCE | Soft guidance, applied when nothing stronger applies |
| 10 | EXAMPLE | Worked examples |
IDENTITY and MISSION compile as prose; the rest compile as structured
lists.
Deterministic compilation
The compiler is a pure function: the same rules always produce the same prompt, byte for byte. Nothing is generated at runtime, so what you preview is exactly what runs.
The compiled prompt includes an explicit precedence instruction — safety and platform policy first, then your rules in section order — so the model has a tiebreaker when two rules pull in different directions.
Live preview
The rules editor shows the compiled prompt as you write it, colour-coded by section, so you can see how your rules assemble into the instruction the model actually receives.
Conflict analysis
Rules accumulate. Rule 3 says "always cite a source"; rule 11, added months later, says "keep answers to one sentence". Nothing errors — the agent just gets worse in ways that are hard to trace.
Conflict analysis is an on-demand check that reads your rule set and flags contradictions. It runs when you click it, never on every keystroke.
Conflict analysis uses your own LLM key (BYOK) — it spends your provider tokens, not platform credits. It's rate-limited to 10 runs per hour.
Inheritance and node overrides
Every LLM node in the agent inherits the compiled rules. A node's own system prompt field becomes an appendix — appended after the shared rules — for instructions that apply only to that step.
This means shared behaviour lives in one place, and node prompts stay short and specific.
Toggling and versioning
Each rule can be enabled or disabled individually without deleting it, and carries its own author and timestamps. Turn a rule off to test its effect, then turn it back on.