$49Prince Maurya · AI engineer
Agents that still work on the eighth run.
I build open-source infrastructure for tool-using LLMs — harnesses, training loops, security gateways, and eval gates — then publish what the benchmarks actually show.
Open source on GitHub · Weekly essays on agent reliability
01 / Selected work
Open source, built to be inspected.
Four systems for the parts of agents that demos skip: permissions, training, tool security, and production policy. All MIT-licensed.
01
Argus
Agent harnessA small, inspectable Python agent harness: path-jailed tools, an approval flow, secret redaction, durable sessions, six providers, and zero runtime dependencies.
Python · v0.2.0 · MIT
02nano-agentrl
GRPO trainingSearch-R1 re-implemented from scratch in PyTorch. Multi-turn GRPO teaches an LLM to reason and call a search tool: GSM8K 32.4% → 47.9%, search-agent EM 0 → 19.6%, plus pass^k reliability analysis.
PyTorch · v0.1.1 · MIT
03MCP Sentinel
Tool securityA runtime security gateway and evaluation bench for MCP and tool-using agents — built around the fact that an approved tool server can change after trust.
Security · MIT
04Agent Reliability OS
ObservabilityProduction-style observability, evaluation, and runtime policy control for tool-using LLM agents: traces, eval gates, and the controls that catch drift before users do.
Infrastructure · MIT
02 / Products
Ready-made systems for Claude Code.
Two ways to skip the setup: a market-research system for Claude Code, and a build path for designers shipping real apps.
03 / Writing
The Reliable Agent
Production systems engineering for AI agents. Weekly benchmarks, eval gates, MCP security, Agentic Memory, and GRPO post-training loops. Free in your inbox, unsubscribe anytime.
Latest essays4 published
01 / ReliabilityPass@8 was 28%. Pass^8 was 4%. Same model.02 / EvalsWhy your agent eval passes locally and fails in production03 / SecurityThe MCP rug pull: your approved tool server can change after trust04 / TestingWhat 102 tests taught me about agent reliabilityRead everything on Substack 04 / About
Built by Prince Maurya.
I'm an AI engineer focused on the unglamorous parts of agent systems: tool permissions, eval gates, training loops, and what happens when a model runs a hundred times instead of once.
Veyra is where I ship the systems that come out of that work: open-source infrastructure on GitHub, Claude Code products for builders, and The Reliable Agent, a weekly letter about production agent engineering.
01
Inspectable by default
Tools should be readable, permissions explicit, and side effects visible. A system you cannot inspect is a system you cannot trust.
02
Evidence over demos
Benchmarks, failure modes, and reproducible runs. A good demo shows the best case; engineering is about the rest.
03
Reliability is repetition
The interesting question is what happens on the eighth run, not the first. Pass^k matters more than a single pass.
01What is Veyra?
Veyra is the studio and product shop of Prince Maurya, an AI engineer working on agent reliability, evaluations, and Claude Code systems.
02Are the open-source projects free?
Yes. Argus, nano-agentrl, MCP Sentinel, and Agent Reliability OS are MIT-licensed and public on GitHub.
03Where do I buy the products?
Village Builder System is sold on Gumroad and Claude Code for Designers is sold on Contra. Checkout, delivery, and refunds are handled by those platforms.
04Do I need to code to use the products?
The Village Builder System requires Claude Code. Claude Code for Designers is written for designers and creators, not only developers.
05Why is the newsletter called The Reliable Agent?
Because shipping an agent once is easy. Making it behave the same way on the hundredth run is the actual engineering problem.
06What do you write about?
Production agent engineering: benchmarks, CI/CD eval gates, MCP security, Agentic Memory, and GRPO post-training loops.
Start here
Build something that survives repetition.
Start with an open-source system, a Claude Code workflow, or the writing behind both.