IBM Agentic AI Dev Day, four-person team

AegisAIgovernance layer for agentic AI

2026. Repository.

What I ownedwatsonx.ai prompt design and the reasoning logic: the RESPOND, ESCALATE and REFUSE decision layer, the policy-confidence model and the rule ordering.

  • Milan Orchestration
  • Mahathi Tools and backend
  • Meenakshi Frontend and docs

Architecture, drawn

AegisAI: every request runs through a priority-ordered rule set, refusal conditions first, and leaves as one of three verdicts. Policy confidence is computed from source authority, freshness and agreement, and low confidence pushes toward escalation.Decision layer, what I ownednononolow confidenceRequestfrom any of 5 agents1. Immediate refusalchecked first2. Informationalanswer directly3. Safe autonomousact without sign-off4. Everything elseescalate to a humanPolicy confidence0.4 authority + 0.3 freshness + 0.3 agreementREFUSERESPONDESCALATE4 governed tools, 3 verified demo scenarioshackathon prototype, not production
AegisAI: every request runs through a priority-ordered rule set, refusal conditions first, and leaves as one of three verdicts. Policy confidence is computed from source authority, freshness and agreement, and low confidence pushes toward escalation.

Problem

An agent that can call tools needs a rule for when it may act on its own and when a human must sign off. AegisAI makes that rule explicit and computed, instead of leaving it to whatever the model feels like in the moment.

Architecture

Every request passes through a decision layer that returns one of three verdicts: RESPOND, ESCALATE or REFUSE. The verdict comes from a priority-ordered rule set: immediate refusal conditions first, then informational requests, then actions safe to take autonomously, then everything else escalates to a human.

Confidence in the governing policy is computed: source authority weighted 0.4, freshness 0.3, agreement between sources 0.3. Low confidence pushes a decision toward escalation.

Three Prompt Lab templates were validated before being operationalised. Around the decision layer, the system runs 5 collaborating agents over 4 governed tools.

Decisions

Refusal is checked first
Rules are ordered by cost of being wrong. A missed refusal is worse than an unnecessary escalation, so refusal conditions run before anything else can answer.
Confidence is a formula, not a feeling
Three weighted, inspectable terms. When the layer escalates, a human can see which term dragged the score down.

Measured, not asserted

Collaborating agents / governed tools
5 / 4
Prompt templates validated before use
3
Demo scenarios verified
3

Honest limits

  • Hackathon prototype, explicitly not production-deployed.
  • No user or revenue metrics exist, and none are claimed.

Stack

  • Prompt design
  • watsonx.ai