AI Research
Sumit P. Ahuja · Main Labs, New York
Thesis
Multi-agent AI is rediscovering coordination and reliability failures that inter-domain networking solved decades ago.
Agents delegate authority across organisational boundaries. They cannot tell a refusal from a failure. They reserve capacity they cannot release. They corroborate a claim with an echo of itself. Routing, congestion control and network management met each of these between 1981 and 2005 and shipped mechanisms that make the internet what it is today.
StandardsIETF, 2026
Contributor across the agentproto, agent2agent, dawn and dmsc efforts, and the AUDIT BoF.
Internet-Draft authored
draft-ahuja-agent-routing-policy, A Policy Grammar for Inter-Domain Agent Routing. Individual submission, Informational. Filed 17 September 2026 and revised since.
Working papers
-
Sumit P. Ahuja
Running large language models across a mix of tiers promises lower cost and latency by resolving easy queries locally and escalating hard ones upstream. That escalation hinges on a confidence signal the small model emits about its own output. This paper shows the binding constraint inverts: where compute is billed, cost decides; where it is sunk on owned hardware, signal quality decides.
-
Sumit P. Ahuja
Agentic workloads run for a long time but sit idle in tool calls for most of it, which makes reserving capacity for them expensive. This paper puts that reservation on a pre-registered test bench, measuring across 108 configurations against six equally tuned baselines whether per-flow completion guarantees buy more completed work per provisioned GPU-hour than they strand. Throughput alone does not carry the reservation, but the guarantees do deliver interactive tail-latency isolation and bounded containment of runaway flows.
-
Sumit P. Ahuja
Agent tasks are increasingly delegated across organizational boundaries, and existing work identifies the absence of a way to express capability policy between domains as a gap. This document defines four policy attributes for inter-domain agent delegation, the declarations each attribute carries, and a validity condition on delegation chains that no single party establishes by observing the whole chain.
-
A Safety Kernel for Escapable AI Systems: Ensuring Alignment at the Execution Layer
Seth Dobrin, Łukasz Chmiel, Sumit P. Ahuja
Controls that live inside an AI system's own runtime, such as system prompts, output filters and guardrail libraries, can be reached by anything that influences that runtime, including the system itself. This paper argues that such escapable systems need alignment enforced at execution time, outside their trust boundary, and states five properties that layer must satisfy, adding revocable compute to four authorization properties. It presents an open-source Rust kernel that realizes all five, and evaluates it in a pre-registered campaign in which the governed system's own self-improvement loop is tasked with defeating it.
StandardsContributions
Cited by two published Internet-Drafts
- Draft
Citation
- draft-konda-agentproto-evaluation-state-02 §4.2, §10, §14
Cites draft-ahuja-agent-routing-policy-02 §4, §6 and §9; acknowledged for the discussion of UNDECIDED reasons, PROFILE_NOT_HELD reports and chain validity at delegation hops
- draft-sergeev-audit-task-requirements-00 §2.10, Acknowledgments
Cites the 1 October agent2agent post on the crossing distinction; acknowledged for expected-record and cross-domain auditing discussions
Credited in IETF working-group discussion
- Source
Credit
- agentproto co-chairs Orie Steele and Leslie Daigle, Topic 1 summary, 19 Oct 2026
Link the 3 October post as "Sumit's contribution" on delegation between independently administered domains
- Lars Kersten Kroehl, agentproto Topic 2 terminology table, 21 Oct 2026
Maps draft-ahuja-agent-routing-policy-03 as one of five terminology sources in Reference Architecture input
- Stelio E. Efstathiou, DMSC list, 27 Oct 2026
Announcing SAMP -04, thanks "Sumit Ahuja and Dave" for comments on late resolution, temporal evidence and event ordering
- Mikhail Sergeev, AUDIT list, 20 Oct 2026
Supports "making the exception in Sumit's wording explicit" in the charter scope
- Girish Konda, agentproto Topic 1, 19 Oct 2026
"Strengthens Sumit's ask" to carry the cross-domain case as its own deployment model
- Jijie Wei, DMSC list, 13 Oct 2026
Cites draft-ahuja-agent-routing-policy on where the containment relation lives: the governing profile
- Chong Feng, agentproto Topic 1, 4 Oct 2026
"+1" to the point that a cross-administrative-domain variant of use case 3.3.2 is missing
- Girish Konda, agentproto WG Review, 22 Sept 2026
Agrees with the 17 September charter comment that revocation is the hard half of the lifecycle
- Gareth Wong, AUDIT list, 17 Sept 2026
Calls the shared absence-of-corroboration point "spot on"
Named in the acknowledgements of eight published Internet-Drafts
- Draft
Contribution
- draft-schrock-ep-bounded-capability-receipts-06 §22
Scope-comparison analysis behind three normative requirements
- draft-bu-agentproto-security-principal-binding-08 §36
Two separate entries
- draft-agentic-ai-usecases-requirements-02
Acknowledgements
- draft-feng-agentproto-session-requirements-01 §8
Session-transfer inheritance ambiguity
- draft-wadkins-agentproto-action-determinability-00
Three contributions
- draft-efstathiou-samp-agent-management-04 §16
Technical review
- draft-bradleyb-audit-decision-records-00
Deadline-derivation disclosure and staleness analysis
Normative text authored and published
Requirement B3-12, merged verbatim into draft-agentic-ai-usecases-requirements-02 (now B3-8 in -05):
The protocol is required to enable a receiving party to determine that the authority conveyed to it does not exceed the authority conveyed to the hop it received from.
A published verification rule at Council of AI on what a verifying signature does not establish, registered as correction C-2026-0902-09, text here.
Measurement
External methodologist on a pre-registered study of delivery incidence in agent commerce. Review appended to the study under name, with an independently authored blind control for the study's grading instrument.