Nikhil Navali
Senior software engineer building agent runtimes for long-horizon autonomous work.
Building autonomous agents has been the most absorbing work of a decade-long career in software — the one where execution stopped being the job and judgement became it. Models are capable enough now that telling one what to do is rarely the hard part. The hard part is everything that has to be true before someone will act on what it produced.
The intelligence comes from the model. The competence comes from the engineering.
Autonomous execution engine for open-ended work
2026Decomposes a goal, selects its own tools, and cites a source for every claim. A separate verifier re-checks the finished work against those sources, with per-deliverable gates flagging anything unsupported before it ships. Each run deposits durable lessons the next one inherits.
~2,000 runs · a day of engineer work returns in under an hourHuman audit removed from a production ML release path
2026Agents now hold the ambiguous precision/recall tail — the cases that previously required a person — across tens of thousands of decisions a day. Accuracy is measured under independent audit rather than self-reported.
95%+ accuracy under independent auditMulti-tenant agentic AI platform
2026Agent runtime, 25+ tools, RAG over domain knowledge bases, an MCP tool stack, and run observability — built and owned end to end. A second team stood up its own domain agent in under a day by contributing only a knowledge base.
100+ users · 50+ autonomous investigations dailyOwn the memory, rent the model
Sep 2026Five months running Karpathy's llm-wiki pattern: what a portable memory layer actually buys, where it's worth diverging from the gist, and the three ways maintenance quietly fails.
Write the standards, not the steps
Sep 2026An autonomous research agent had to move to a domain it was never built for. The port took two days — and the reason was a ratio: nine parts standards, one part domain.
- Agents
- agent runtime designlong-horizon autonomyplanner/executor orchestrationtool use & function callingMCPcontext engineering
- Evaluation
- LLM-as-judge designeval frameworks & golden setsindependent verificationcitation grounding
- Systems
- PythonScalaSparkAWSdistributed data at scale