Reliable agent execution
Typed state, model routing, tool permissions, recovery, verification, and human approvals for agents that must do more than generate text.
Research output · Agent Runtime + Agent Studio
Applied research & development
Algen is an AI product and applied research company. We study how agents are built, evaluated, governed, discovered, and trusted—then turn useful findings into open-source infrastructure, commercial products, and production systems.
Publications
arXiv preprint · Oct 2026
arXiv:2610.18603
cs.AI · cs.CY · eess.SY
Nutan Kumar Naik, Vijay Prasad Poudel, Aditya Kumar Saroj, Saurav Samantray, Abhishek Patel
A maturity model that grades AI agents from 0 to 7 stripes across 12 areas, based on how strong the evidence is. We tried it on 30 public Python agent repos. The first results are exploratory.
Read on arXivUnder review · Sep 2026
IEEE Transactions on Dependable and Secure Computing
Nutan Kumar Naik, Aditya Kumar Saroj, Vijay Prasad Poudel, Saurav Samantray, Abhishek Patel, Sambhav Jain, Krish Raj
Where in an agent's run should safety rules be checked? 3,000 trials across LangGraph, CrewAI and OpenAI Agents compare five placements. Checking actions before they run cut unsafe actions from 90% to 26%, and adding monitoring caught every one that still got through.
Read the preprint (PDF)Submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible.
arXiv preprint · Jul 2026
arXiv:2607.14309
cs.AI · cs.CY
Nutan Kumar Naik, Aditya Kumar Saroj, Vijay Prasad Poudel, Saurav Samantray, Abhishek Patel
A governance stack on OpenTelemetry for alignment drift and shadow AI, producing tamper-resistant evidence packages mapped to EU AI Act requirements, without exposing private data.
Read on arXivResearch focus
Typed state, model routing, tool permissions, recovery, verification, and human approvals for agents that must do more than generate text.
Research output · Agent Runtime + Agent Studio
OpenTelemetry-native traces, datasets, scorers, policy checks, incidents, and evidence for understanding and controlling agent behavior.
Research output · Traccia
Versioned manifests, maturity, risk, permissions, safeguards, and evidence that help teams assess an agent before adoption.
Research output · Agent Hub + ASCEND
How agents, existing software, operational data, and people can share decisions with clear authority, explanations, and measurable outcomes.
Research output · DIRECT method + production systems
How we publish
We separate documented facts from hypotheses, describe methods and limitations, and distinguish our own experiments from independent benchmarks. Research is useful when another engineer can inspect the reasoning and apply it responsibly.
Inspect and contribute to the foundations behind Algen agent systems.
Explore Agent RuntimeSee how research questions become software for building and operating agents.
View Algen productsRead technical analysis, experiments, benchmarks, and implementation lessons.
Read the research notesSee how ideas are tested inside real workflows and operational constraints.
View case studiesRecent research notes
AI Engineering · 2026-11-11
Your primary model provider starts returning 503s. Retry harder and every user waits; fall back on any error and your bugs hide behind a weaker model. A hands-on tour of how the runtime routes around an outage, with a script that breaks the primary eleven ways.
AI Engineering · 2026-11-07
A refund times out. Retry and the customer may be paid twice; don't and they may not be paid at all. How the runtime keys, records and reconciles every tool call so it takes effect once.
AI Engineering · 2026-11-04
One user request can become a dozen model calls, most of them unseen. How the runtime reserves, settles and releases every call against the right account, and stops a run cleanly when the credits run out.