Governing Autonomy: Agentic AI, Multi-Agent Systems, and the Infrastructure of Trust

Abstract

The session examined the emerging governance gaps in agentic and multi‑agent AI systems. After a short introductory framing, the panel contrasted traditional AI with the broader spectrum of autonomous agents, highlighted the need for run‑time governance, and discussed certification, standards, and a five‑layer “governance stack.” A live demo of Cognizant’s open‑source Neuro AI Multi‑Agent Accelerator illustrated practical orchestration, observability, and trust mechanisms. The discussion broadened to analogies from safety‑critical domains (aviation, drones), the balance between engineering best‑practices and regulatory frameworks, economic implications for startups, and the importance of auditability, open standards, and a commons‑oriented approach to trustworthy agentic AI.

Detailed Summary

Mahesh (Project Nanda) opened the session, positioning his team as builders of the “Internet of AI agents” and emphasizing that trust must be baked into the infrastructure from the first principle, not added as an after‑thought.
He invited the panel to reflect on the autonomy‑fear gap – many attendees had either built an agent or felt uneasy about agents, but most (≈80‑90 %) did not consider agents dangerous.

2. What Is an “Agentic” System? – Spectrum of Autonomy

Ellie Sakhaee (Google) clarified that agentic AI is not a binary label; it exists on a continuum ranging from simple chat‑bots (e.g., Google Deep Research) to fully autonomous vehicles.
Key dimensions defining the continuum: memory, short‑term vs. long‑term planning, execution capability, and degree of real‑world action.
She warned that because agents become network actors (visible or invisible), runtime governance is mandatory; traditional pre‑deployment checks are insufficient.

3. Governance Gaps in the Current Landscape

Ellie traced the evolution of governance models:
1. Machine‑to‑Machine → early SaaS → micro‑services → today’s agent‑to‑agent interactions.
2. Existing governance assumes a central authority that enforces policies after scale is reached. With autonomous agents operating at scale, that assumption breaks down.
Two core challenges highlighted:
1. Safety & Trust – ensuring agents do not act unexpectedly.
2. Accountability & Openness – preserving transparency while allowing open, decentralized deployment.

4. Certifying Agents – The IEEE Perspective

Alpesh Shah (IEEE) argued that certification requires a clear definition of what good looks like for an autonomous system.
He emphasized iterative governance: continuous monitoring, contextualizing triggers, and maintaining transparent, accountable logging.
IEEE has already contributed to standards covering data transparency, age‑appropriate design, accountability, and privacy for opaque AI systems.
The standards development process involves a broad multidisciplinary community (lawyers, doctors, artists, engineers) to capture diverse risk perspectives.

5. Conceptualising a “Governance Stack”

Alpesh introduced a five‑layer “5‑K‑Clear” model for autonomous AI:
1. Build‑time – data governance, model versioning, provenance.
2. Deploy‑time – policy tracking, permissioning, secret management.
3. Runtime – real‑time observability, kill‑switch capability.
4. Remediation – audit trails, incident‑response architecture.
5. Accountability – reporting structures, post‑mortem analysis, compliance mapping.
He stressed that embedding governance into the product (not as an afterthought) becomes a competitive GTM advantage; enterprises now demand auditability, traceability, kill‑switches, and cost‑of‑failure transparency before signing contracts.

6. Start‑ups & the “Minimal Viable Trust Stack”

Alpesh proposed a four‑component minimal stack for early‑stage agentic AI companies:
1. Agent Identity Registry – immutable metadata (origin, training data, ownership).
2. Orchestration Guardrails – real‑time policy enforcement at the orchestration layer.
3. Observability Architecture – fine‑grained logging (token usage, cost, ESG metrics).
4. Defined Oversight – clear human‑in‑the‑loop or human‑in‑command controls.
Without these, launching an agentic ecosystem in production is deemed high‑risk.

7. Live Demonstration: Neuro AI Multi‑Agent Accelerator

Praveen Tanguturi (Cognizant AI Lab) showcased the Neuro AI Multi‑Agent Accelerator, an open‑source, low‑code/no‑code framework for building and scaling multi‑agent networks.
- Key Features (as demonstrated):
  - LLM‑agnostic & Cloud‑agnostic – users choose the underlying large language model and deployment environment.
  - Protocol Interoperability – supports OSR, A2N, MCP server protocols for agent‑to‑agent communication.
  - Built‑in Security & Guardrails – security is not an after‑thought.
  - Graphical Orchestration UI – even non‑technical stakeholders can assemble agent networks in minutes.
  - Dynamic Agent Creation – audience prompt (“Create a multi‑agent network for India AI Summit”) resulted in instant generation of agents, role definitions, and connection to an orchestrator.
  - Observability & Auditing – real‑time chat view of inter‑agent messages, token‑level cost breakdown, ESG scoring (energy usage, carbon footprint).
The demo highlighted reuse of existing leaf‑node agents (orange icons) to avoid redundant creation, illustrating the framework’s catalogue‑driven composability.
Audience interaction: a question about managing heterogeneous contexts/memories across concurrently running agents was acknowledged but deferred for a deeper offline discussion.

8. Cross‑Industry Safety Analogies

Ellie drew parallels with aviation and drone regulation:
- Early drone rules required Visual Line‑of‑Sight (VLOS) – a human pilot constantly supervising.
- As AI‑driven detect‑and‑avoid capabilities matured, regulators moved toward Beyond VLOS, trusting the AI system more, while keeping a human‑in‑command role for higher‑level oversight.
The panel concluded that agentic AI will follow a similar trajectory: start with human‑in‑the‑loop and gradually shift to human‑in‑command as safety mechanisms prove reliable.

9. Balancing Engineering Practices vs. Regulation

Mahesh (Project Nanda) argued for parallel development of innovation (the data plane) and governance (the control plane). Heavy‑handed compliance early on could stifle innovation; governance must evolve alongside.
Amir Banifatemi (Cognizant) added that early‑stage experimentation benefits from framing (guidelines) rather than strict regulation. Once a technology matures, formal standards become valuable for market acceptance and differentiation.
Apoorva Goyal (Insight Partners) highlighted the economic tension:
- Seed/Series‑A startups selling into enterprises already need enterprise‑grade governance (SOC‑2‑style, auditability).
- The most successful AI‑native teams treat governance as a weekly operational cadence – continuous evaluation, red‑team exercises, post‑mortem analysis.
He also referenced India’s digital infrastructure success (Aadhaar, UPI, DigiLocker) as a model where government‑mandated standards enabled a vibrant ecosystem of private‑sector innovation on top of shared foundational layers (identity, payments, privacy).

10. Infrastructure for Trust – Auditability & Collective Learning

Ramesh Raskar emphasized the need for immutable audit trails and open‑source platforms that enable anonymized sharing of failure modes and performance metrics across organizations.
A global commons for trustworthy AI (e.g., AI Commons led by Amir) was proposed as a mechanism to:
- Treat foundational AI technologies as public utilities.
- Democratize access to safety tooling, standards, and evaluation benchmarks.
- Reduce jargon and make concepts accessible to non‑technical stakeholders.
Ali (likely Ramesh) suggested that benchmarking multi‑agent systems (as discussed in the Distributional AGI Safety paper) is essential. Labs should report transparent test results against such benchmarks to build confidence.

11. Open Questions & Future Directions

Open Question / Debate	Summary of Positions
How to manage heterogeneous LLMs across agents?	No consensus reached; participants agreed further research and standards are needed.
When does regulation become a competitive advantage vs. a cost?	Consensus: early integration of governance becomes a differentiator for enterprise sales; for PLG‑focused consumer products, governance may be deferred.
What single “infrastructure” piece can drive progress?	Multiple viewpoints: a) Robust evaluation benchmarks; b) Immutable auditability layers; c) Public‑utility style commons. No single answer, but all agree on the need for collective, open frameworks.

12. Closing Remarks & Call to Action

Amir Banifatemi reiterated the importance of AI Commons as a “public utility” for agentic AI, stressing that education, demystification, and open standards are essential to avoid public fear and foster responsible adoption.
The panel thanked the audience, encouraged participants to explore the Neuro AI Multi‑Agent Accelerator GitHub repo (Apache 2.0 license), and invited continued dialogue on governance, standards, and trust infrastructure.

Key Takeaways

Agentic AI lives on a continuum of autonomy; governance must therefore be runtime‑focused, not only pre‑deployment.
IEEE standards are already addressing data transparency, age‑appropriate design, accountability, and privacy for autonomous systems.
A five‑layer governance stack (build, deploy, runtime, remediation, accountability) provides a holistic blueprint for trustworthy agentic AI.
Start‑ups should implement a Minimal Viable Trust Stack (identity registry, orchestration guardrails, observability, oversight) before production release.
Neuro AI Multi‑Agent Accelerator demonstrates that low‑code, open‑source tooling can embed security, observability, and ESG metrics directly into agent networks.
Safety analogies from aviation/drones suggest a future shift from human‑in‑the‑loop to human‑in‑command as AI safety matures.
Balancing engineering and regulation requires parallel development of innovation (data plane) and governance (control plane); heavy early regulation can hinder progress, but standards become critical at scale.
Governance is becoming a GTM advantage: enterprise buyers now demand auditability, kill‑switches, and traceability before signing contracts.
Auditability, immutable logs, and global benchmark sharing are the most promising levers for collective learning and trust building.
AI Commons and public‑utility thinking aim to democratize safe agentic AI, reduce jargon, and ensure inclusive access to foundational capabilities.

Prepared from the verbatim transcript of the Delhi AI Summit session “Governing Autonomy: Agentic AI, Multi‑Agent Systems, and the Infrastructure of Trust.”

See Also:

India AI Impact Summit 2026

Explorer

governing-autonomy-agentic-ai-multi-agent-systems-and-the-infrastructure-of-trust