The OPERA loop
Seven phases — in order, every time
O — Observe
Read data via the adapters; compute data confidence. Below 0.60 → flag, continue with a warning.
P — Plan
Pattern-match first — a predefined plan on a hit. Otherwise the LLM plans, but only on abstract structure — not raw values.
E — Execute
Runs only registered capabilities. Safe read and analysis steps retry on failure (up to 2) with a per-step timeout; an external write is never retried automatically — a failed write goes to a person to decide.
R — Reason
The LLM interprets the findings — statistics and counts only, max 200 tokens, a strictly bounded prompt.
A — Act
Build the structured output and enforce Culture Rules — impact quantified in EUR, exactly one recommendation. Nothing is delivered yet.
V — Verify
A fail-closed gate before anything ships: the Dual-Brain Critic and End-State Evaluator review the output, and RAL proves every critical value against its source. Unproven or unsafe output is blocked, never delivered.
D — Deliver
Only verified output is delivered per tier — READ_ONLY auto-delivered, PROPOSE_CHANGE queued for a named approver — a verified report produced for your review and every step logged.
The Reason phase receives statistics — counts, trends, thresholds — not raw amounts or names. This is ABN's No-Data design: raw operational data is not sent to the prompt during normal operation.
Blueprint Generator
Every agent is unique per (customer × process)
ABN is not a catalogue of fixed agents. Blueprint Generator creates a signed, unique blueprint for every process discovered — calibrated to your industry and your own baselines.
DNA Phase
Read process graphs, detect the industry, and compute your own baselines — your actual deviation level, not a generic industry default.
Genesis prompt
The LLM receives only abstract structure + your terminology + your baseline. Generates the agent name, goals and a recommendation template — not raw values.
Validation + HMAC
AVM reviews against industry rules. Culture Rules force a EUR figure and one recommendation. Then HMAC signing. The blueprint is locked.
Agent archetypes
Four active archetypes in V1
MONITOR
Monitors thresholds and alerts. No analysis — just "is this over the limit?". Fastest and cheapest.
ANALYZER
Looks for deviations between sources. RAL-attests every finding and quantifies the impact in EUR.
PLANNER
Proposes schedule changes and actions. Generates proposals — never decisions, always for approval.
COORDINATOR
Orchestrates multiple agents via the Case object with scope protection. Born only when there are ≥ 3 independent sub-tasks.
EXECUTOR is a fifth archetype that exists in the code but is locked until V2 — gated by a separate release decision and future approval, policy and security controls.
Never without your approval
The human has the final say — always.
Tiers are hardcoded in the Commitment Gate. Nothing can break past them — not even an LLM that hallucinates confidence.
Tier 1 · READ_ONLY
Auto-delivery
Read-only, no impact. Delivered automatically at high confidence.
Tier 2 · PROPOSE_CHANGE
Proposes — you approve
Saves the proposal to the queue. A named employee must approve before anything happens.
Tier 3 · EXECUTE_CHANGE
Locked in V1
Direct writes to source systems. Off until the required approval, policy and security controls are in place and you actively enable it.