Agents that can act — with an architecture that can stop them.

Guild AEGIS and ControlPlane

AEGIS is the operating capability: intelligent agents, orchestration, runtime controls, data and process quality, evidence, cost management, and measurable business outcomes. ControlPlane is the connected view a CIO can actually use — not another AI dashboard.

• Inventory — approved, experimental, embedded, third-party, and shadow AI
• Accountability — named owners, recertification, exceptions, retirement
• Lineage — data, models, prompts, tools, APIs, and downstream systems
• Control — identity, policy, human approval, and a tested kill switch
• Cost — models, tokens, infrastructure, oversight, and compliance
• Evidence and value — what can be proven, not what was deployed

Three layers, or it is not enterprise-ready

We treat agentic AI as architecture, not a chatbot project.

• Agent runtime — perceive, reason, plan, remember, act, validate, respond
• Enterprise control plane — identity, registry, entitlements, runtime policy, human approval, kill switch
• Observability and evidence — who acted, on what data, with which tool, under which policy, with what outcome

Controls intercept execution

An agent should not execute a sensitive action and have governance discover it in a log. The path is Agent → Identity → Authority → Context → Policy → Risk → Human approval when required → Action. That is what watsonx.governance and OpenPages are for when they are wired into the runtime, not bolted onto a wiki.

Where we put agents to work

Every pattern includes a human in the loop where the risk requires it.

• Incident response — Instana detects, the agent diagnoses, a playbook remediates, MTTR drops
• Regulatory change — monitor, assess impact, draft updates for human review
• Procure-to-pay — validate budget and policy, route approvals, talk to approved vendors
• Model risk — continuous bias, drift, and performance monitoring
• Finance ops — reconcile, flag anomalies, prepare journal entries with an audit trail