Most AI agents are a black box.
Here are all 19 of ours.
LiminaHub's Observe engine is 19 real agents across 6 layers — each mapped to the code that runs it. The governing rule: the model reasons; deterministic code enforces every guarantee.
A language model decides the approach — how to interpret your goal, what to look at, what matters. Swap in a weaker model and the behaviour visibly changes, because the model is genuinely doing the thinking.
Every number you trust, every permission, every audit entry is deterministic code — never the model's discretion. Swap the model and those stay byte-identical. That line is what makes an AI agent safe to run your business.
The 6 layers
On a single Plan & run, 13 run-path agents engage; the rest are the always-on ingestion and reliability crew.
Orchestration & Runtime
4 agentsDecides what runs, in what order, within budget.
Owns one objective run end-to-end — decides which agents run each pass and sequences the loop, without ever calling a tool itself.
Decides which objective's pass runs next, at what concurrency and budget, with fair-share so one objective can't starve the others.
Runs a single step's LLM loop — the only tool-caller — invoking just the bound read-only tools through the policy gate.
Keeps runs honest: single-flight, idempotency, cost and iteration ceilings, orphaned-run reaping, and crash recovery by idempotent re-run.
Objective Understanding
3 agentsTurns your plain-English goal into a precise, grounded spec.
Turns a free-text objective into a structured, grounded watch spec — KPI, observables, plan, and clarifying questions.
Binds each observable to a concrete read-only tool action, validates it through the policy gate, and records any gaps.
Computes each objective's KPI time series and learns its baselines and deviation bands.
Event Ingestion
3 agentsThe read-only sensory edge — real-time and batch, always on.
The read-only sensory edge — both scheduled polls and real-time per-provider webhooks (verified, captured, real-time wakeup), with full provenance.
Turns raw signals into canonical events (gmail → email.received) and resolves the sender into a stable entity — many emails from one person collapse to one.
Fans each canonical event out only to the objectives whose subscription actually matches.
Analysis & Monitoring
2 agentsRanks what matters and answers questions — grounded, never invented.
Ranks open concerns by impact and calibrated confidence (high / medium / low), each with a one-line 'likely because'.
Ask anything over your data, plus the scheduled daily digest — grounded strictly in citable rows, never invented.
Shared Context & Memory
3 agentsOne consistent belief per objective, plus durable, cited recall.
The single writer of an objective's shared belief and append-only context log — with a durable, always-on listener that wakes objectives in real time.
Distills observation clusters into compact insights, then embeds and indexes them for recall.
Retrieves the most relevant durable memory for a question — workspace-scoped, with citations; it never generates.
Governance & Ops
4 agentsThe deterministic rails that make every guarantee hold.
Classifies every tool call (read / draft / run), hard-pins Observe to read-only, fail-closed — plus PII masking before anything leaves.
Owns the append-only, tamper-evident audit hash-chain — with a verify operation that detects any retro-edit.
Meters every billable op exactly once and enforces budgets and rate limits before expensive work, fail-closed.
Wraps each run in a span tree — latency, tokens, cost, model — keyed to objective and run, and vendor-swappable.
The rails that never bend
These aren't prompts you hope the model follows — they're deterministic code that holds no matter which model is running.
A policy gate classifies every tool call and hard-pins Observe to read-only. If it's unsure, it stops.
Every action is written to an append-only hash-chain — a verify step catches any retro-edit.
Budgets and rate limits are enforced before expensive work runs, fail-closed.
Sensitive data is masked before it reaches a model, a trace, or memory — by default.
This is why an AI agent for your small business can be trusted with real access — see it level by level.
Your Copilot is waiting on your desktop.
Download LiminaHub, start at Observe for free, and turn the dial whenever you're ready.