Tisha Chawla is a Microsoft software engineer and co-creator of AgentPlane, an open-source initiative building infrastructure for reliable AI agents. Her work addresses two critical production failures: agents making consequential decisions that engineers cannot reproduce, and autonomous workflows accumulating costs beyond their operators’ control.
Chawla studied at Vellore Institute of Technology from 2020 to 2024 and interned at Microsoft before joining as a software engineer. Her early projects included a Microsoft Engage 2022 application and a tool for analyzing SEC EDGAR filings. At Microsoft, she has worked on enterprise software-engineering agents, site-reliability workflows, and specification-driven development associated with the Commerce and Ecosystem Data Platform.
Production systems that remain accountable
- Replayability over model determinism. Chawla distinguishes forcing a model to repeat itself from reconstructing an execution that already happened. Request batching, mixture-of-experts routing, floating-point computation, retrieval changes, and tool behavior can produce different outcomes from the same prompt, even at zero temperature. Her work on production-agent failures emphasizes preserving execution state instead of assuming reproducible generation.
- Chronicle and boundary-level recording. With Susheem Koul, she created Chronicle, which records inputs, outputs, and metadata around model calls, retrieval steps, and tools. Their stock-trading example follows an agent that mistakes a requested dollar amount for a share quantity: the API succeeds, but the transaction is wrong. Replaying the captured trace lets engineers stub model responses, exercise repaired safeguards, and turn the incident into a repeatable regression test.
- Run-scoped token governance. Their TokenOps project tracks costs across an entire agent execution and enforces shared budgets while work remains in progress. Interventions include switching models, restricting subsequent calls, and stopping runaway loops. A Microsoft-published analysis reported approximately 79% lower average spending and improved within-budget completion across 27 scored trials; those results apply to the published benchmark, not every production deployment.
Chawla’s engineering priorities place deterministic safeguards around probabilistic models: execution traces, tool-level assertions, spending policies, and circuit breakers govern consequential actions, while subjective response quality requires separate behavioral evaluation.
