Susheem Koul

Susheem Koul is a senior software engineer at Microsoft and a co-creator of AgentPlane, an open-source initiative building practical infrastructure for AI agents. With Tisha Chawla, he develops tools that make agent failures reproducible and impose enforceable spending limits on autonomous workflows.

Koul studied at Birla Institute of Technology and Science, Pilani, from 2015 to 2019. His earlier technical writing examined Git and version-control architecture; more recently, he has concentrated on the reliability, evaluation, and operating costs of systems that place probabilistic models in charge of consequential decisions.

At AI Engineer World’s Fair 2026, Koul and Chawla illustrated the problem with a trading agent that mistakes a requested dollar amount for a number of shares. The resulting order succeeds technically while producing the wrong outcome, leaving conventional infrastructure monitoring blind to the failure. Their conference presentation introduced a more useful operational distinction: engineers need replayable agent executions, not impossible guarantees that hosted models will always produce identical outputs.

• Chronicle: production incidents as regression tests. Chronicle records inputs, outputs, and runtime metadata around model calls, retrieval, and tools. Developers can replay a failed execution, substitute recorded model responses, rerun a modified guardrail, and verify that the original mistake is blocked.

• TokenOps: run-level token governance. TokenOps tracks spending across collaborating agents, tools, and model calls through a shared ledger and control plane. Its policies can intervene before a workflow makes another expensive request.

• Evaluation beyond deterministic assertions. Koul distinguishes conventional software tests from behavioral agent evaluation: tools and safety boundaries require firm assertions, while open-ended agent behavior calls for assessments of grounding, task completion, decision trajectories, reliability, and cost.

• Efficient context and interoperable agents. His writing on prompt caching examines how stable prompt prefixes reduce repeated computation. He also advocates standardized agent-to-agent communication to help specialized agents collaborate without bespoke integrations.