How to Build a System with True Self‑Decision Capability

Introduction

After years of adding "smarter" modules such as recommendation engines, intelligent客服, and risk‑control models, the author’s new Q2 mission is to give the core business system true self‑decision capability—telling the system the goal, not the step‑by‑step instructions.

What Self‑Decision Is (and Isn’t)

Self‑decision means the system, given business goals and constraints, can autonomously perceive environment changes, reason a suitable strategy, execute actions, and continuously refine based on feedback, without human‑in‑the‑loop steps. It is not a sci‑fi general AI, nor a rule engine wrapped in an AI shell, nor an unsupervised system that operates without human oversight.

Example: In the e‑commerce fulfillment system, instead of operators manually adjusting warehouse allocation when a warehouse overloads, the self‑decision system should detect the overload trend, compute the optimal allocation, issue the adjustment, and monitor the outcome, leaving humans only to audit or intervene when necessary.

Overall Architecture

The system is split into four layers—Perception, Reasoning, Execution, and Governance—connected by a data flow that moves context from perception to decision and actions back to monitoring. Governance is a cross‑cutting constraint framework rather than a post‑hoc patch.

Perception Layer

Real‑time event stream integration : All key business events are ingested via Kafka into a unified bus. Flink enriches each event with semantic tags, causal markers, and time‑window context, aggregating raw events into higher‑level situations such as an "warehouse overload warning".

Metrics and knowledge‑graph fusion : VictoriaMetrics stores time‑series metrics, while Neo4j maintains a business knowledge graph linking entities like warehouse capacity, logistics routes, and product turnover days, enabling flexible impact analysis.

External signal ingestion : Weather forecasts, traffic conditions, and supplier status are fetched through an API gateway, standardized, and injected into the event stream.

Reasoning Layer

The brain of the system uses a multi‑agent negotiation architecture instead of a single monolithic model.

Orchestrator Agent : Built on Claude Opus, it understands the current situation, decomposes problems, distributes tasks to expert agents, and aggregates their suggestions into a final decision. Claude was chosen for its leading long‑context and structured‑reasoning abilities and for Anthropic’s mature Managed Agents API (2026).

Domain Expert Agents : Vertical models for specific domains, e.g., a logistics optimization agent that embeds an OR‑Tools solver and a inventory‑forecast agent running a custom time‑series model. These agents use the most appropriate technique—mathematical programming or statistical modeling—rather than a generic LLM.

Arbitration Agent : When expert agents conflict (e.g., logistics wants fewer routes, cost wants lower capacity), this agent applies a pre‑defined business priority matrix and current constraints to select the highest‑utility combination.

Memory Module : Short‑term memory stored in Redis holds the current decision session context; long‑term memory in Qdrant stores historical decision cases with outcomes. Before a new decision, similar past cases are retrieved and injected into the reasoning context, giving the system "learning from experience".

A diagram (included in the original article) illustrates the internal negotiation flow.

Execution Layer

The reasoning layer outputs a structured decision in a strict JSON‑Schema format. Execution follows four safety gates:

Sandbox rehearsal : The decision is first simulated in a mirrored environment using a digital twin built from a real‑time production data snapshot. If simulated deviation exceeds a threshold, the decision is blocked.

Canary rollout : Decisions are released via an OpenFeature‑based feature‑flag system to a small subset (e.g., 5% of orders). After 30 minutes of stable key metrics, the rollout expands.

Real‑time monitoring & automatic rollback : Continuous metric monitoring triggers an automatic rollback to the last known healthy snapshot if anomalies (e.g., spike in complaints or delivery delays) appear.

Manual kill switch : An on‑call dashboard provides a prominent red button that instantly freezes all ongoing autonomous decisions and reverts control to humans.

Governance & Observability

Permission fences : Each agent has a strict OPA‑defined permission boundary (e.g., the logistics agent can only modify routing, never price or inventory). Changes require dual‑approval.

Decision trace chain : Every decision generates a structured OpenTelemetry span enriched with reasoning basis, confidence score, and rejected alternatives, enabling precise post‑mortem analysis.

Explainability panel : Business stakeholders view a natural‑language summary generated by Claude Sonnet that narrates "what the system saw → what it thought → what it did → the result".

Pitfalls Encountered

Dead‑loop between agents : Early versions deadlocked when agents waited for each other's conclusions. Added timeouts and a "fallback conservative decision" policy to break the loop.

Memory hallucination contamination : Bad historical cases polluted new decisions. Introduced result annotations and weight decay; ineffective cases lose weight over time and are eventually purged.

Over‑confident decisions : LLMs tend to output confident answers even when uncertain. Required each decision to include a confidence score and uncertainty interval; low‑confidence decisions are routed to human approval.

Latency budget explosion : Full multi‑agent negotiation plus sandbox and canary checks pushed latency to minutes. Optimized by routing latency‑critical scenarios (e.g., real‑time risk control) through a fast‑path single‑agent + rule engine, and made sandbox rehearsal asynchronous.

Conclusion

After six months, the author notes that the technical barrier to self‑decision is lower than expected; the real challenge lies in organizational trust and process change. The 2026 infrastructure—mature multimodal models, native agent frameworks, and robust observability toolchains—supports the vision, but widespread adoption requires transparent governance, reliable fallback mechanisms, and a portfolio of proven positive cases. Starting with low‑risk scenarios (warehouse sorting priority) yielded a three‑month data set showing the autonomous decision share rising from 0% to ~35% and response speed improving by an order of magnitude while maintaining or slightly improving business metrics.