Research

The formal treatment, argued in public.

Where the firm’s thinking is written down: research-grade papers on identity and access architecture for the agentic era — standards-honest, vendor-neutral, with running code behind every claim.

Runtime Authorization for AI Agents

A Task-Scoped Credential-Brokering Architecture for Humans, Machines, and Agents

Jairo Cadena · Version 1.0 · July 2026 · 13 sections

Authentication at the perimeter fails for non-deterministic actors; authorization must move to runtime. The paper models an agent’s identity as a composite of four claims, organizes authorization into three layers — may-ever, may-now, may-here — and defines the effective grant as task ∩ ceiling ∩ delegator, attenuating at every delegation hop. It specifies the broker that enforces the model, walks an open reference implementation, maps the standards honestly, and closes with the convergence claim: built for agents, these primitives become one access model for everything that acts.

Read the paper · PDF Reference implementation

Cite as

Cadena, J. (2026). Runtime Authorization for AI Agents: A Task-Scoped Credential-Brokering Architecture for Humans, Machines, and Agents (Version 1.0). Cadena Core Group. cadenacoregroup.com/research · DOI: 10.5281/zenodo.21585832

The paper’s serialized treatment runs as the Agentic Identity series — one section at a time, argued with the people building this — on LinkedIn. Version 1.0 is frozen; corrections and extensions appear as dated errata and future numbered versions.