Akshay Shinde / Field notes
Technical notes on AI agents, software systems, and the engineering decisions that make products dependable.
17 ARTICLES / 29 TOPICS
A close reading of TypeSafe AI's Jev: state and typed questions in, calibrated Choice, Score, and Noul answers out. What is documented, what is vendor claim, where it fits in an agent, where it does not, and why it complements LLMs rather than replacing them.
October 8, 2026 /19 min read
An agent is defined by where control flow lives, not by how capable the model is. State and context, planning, observation, memory, verification, recovery, and permissions, and why reliability comes from the software around the model.
October 8, 2026 /16 min read
A repeatable sequence for designing a system: requirements, rough numbers, interfaces and data, the hard part, failure analysis, and the tradeoffs you are choosing. Worked through on a webhook delivery service.
October 4, 2026 /7 min read
Every remote call has three outcomes, not two. Idempotency, retry budgets, deadlines, quorums, and fencing tokens are what you build once you accept that.
October 4, 2026 /9 min read
Once an agent workflow has branches, parallel steps, approvals, and revision cycles, it is a graph whether or not you draw it. Typed state, nodes, conditional edges, reducers, checkpoints, and how to test and version the result.
October 4, 2026 /6 min read
Much of what users experience as agent behavior is decided outside the model: what context it sees, which tools exist, what they are allowed to do, where they run, and how runs are recorded and judged. A practical guide to building that harness.
October 4, 2026 /6 min read
Request/response, queues, logs, webhooks, and streaming connections each trade a different kind of coupling. A practical guide to choosing between them and to the delivery guarantees each one actually gives you.
October 4, 2026 /8 min read
Prefill and decode, time to first token and time per token, the KV cache, continuous batching, prefix caching, quantization, and speculative decoding, plus what application teams calling a model API can control themselves.
October 4, 2026 /8 min read
Agents are nested loops running at different time scales. Each needs a termination contract, a progress signal, a verifier outside the model, and a deliberate way to give up. Budgets, no-progress detection, feedback, and the offline improvement loop.
October 4, 2026 /7 min read
Microservices buy independent deployment and ownership at the price of a network, distributed data, and a platform to run it all. How to draw boundaries, keep data consistent without distributed transactions, and recognize a distributed monolith.
October 4, 2026 /6 min read
Architecture is less about diagrams than about which decisions are hard to reverse, which qualities the system must protect, and how to keep both visible. Quality attributes, C4 views, dependency direction, ADRs, and fitness functions.
October 4, 2026 /6 min read
A monolith is a deployment choice, not a code-quality verdict. How to structure one with enforced module boundaries, scale it further than expected, and extract services later only where it pays.
October 4, 2026 /5 min read
A practical model for bounded agent execution: make decisions, tool calls, state transitions, and stop conditions explicit in application code.
October 1, 2026 /3 min read
Tool schemas are only one part of the boundary. Authorization, validation, side effects, idempotency, and result shaping belong to the application.
September 22, 2026 /3 min read
Learn how to build modern web applications with Next.js 15, including the new App Router, Server Components, and performance optimizations.
December 15, 2024 /3 min read
Master advanced React performance optimization techniques including memoization, code splitting, and virtual scrolling for lightning-fast applications.
December 10, 2024 /7 min read
Learn essential TypeScript patterns, advanced types, and best practices for building maintainable and scalable applications.
December 5, 2024 /8 min read
Writing
Technical notes on AI agents, software systems, and the engineering decisions that make products dependable.
