// writing
Ideas argued out in public — mostly the ones behind the projects: verifying agents instead of trusting them, tracing code through indirection, and keeping systems replayable.
The best AI tools ship terminal UIs, and suddenly everyone wants one. What .NET offers (Spectre.Console, praised and then audited), what Ink and Bubble Tea do differently, why Go keeps winning this niche — and the pattern I actually ship: a Go TUI attached to a .NET backend over SSE, with a working example.
Event-driven systems rot into callback soup where no two runs behave alike. Here's the cure I used in a trading engine: one queue, one pure reducer, one journal, one effect executor — and the discipline that makes two replays byte-identical.
Two processes, one firehose of events, no broker. What ZeroMQ actually is, which socket pattern fits which job, real throughput numbers, the foot-guns that eat your messages silently — and the lock-step protocol that buys determinism over a socket.
Coding agents will work unattended for hours — if you stop believing anything they say. The full anatomy of an autonomous delivery loop: agent CLIs as processes, a ten-outcome vocabulary, independent gate batteries, fix sessions armed with evidence, and state that survives a power cut.
Most backtesters are a second implementation of the strategy — so they prove nothing about the code that actually trades. Build the engine as a pure reducer and the backtest, the live path, and the debugger become the same thing.
The first version of my orchestrator judged agent sessions as success or failure. Reality needed eight outcomes — and the distinction that matters most is whose fault the failure was.
In a mediator-and-DI codebase, nothing calls anything — everything is connected by convention. Honest tracing means building edges from joins, and making every edge cite the file and line it came from.
Agents burn tokens re-discovering the same codebase every session. Here's the case for handing them a structured map instead — and how DevContext builds one with Roslyn.
Layered/onion architecture optimises for a kind of reuse most services never need. Feature-first vertical slices optimise for the thing you actually do all day: change one behaviour without touching five folders.
Handing an agent a shell is easy and, past a toy, a mistake. The reliable path is a small set of typed, validated, least-privilege tools it can't misuse — which is an API design problem, not a prompting one.
Autonomous coding agents lie by omission: they report success they didn't earn. The fix isn't a better prompt — it's an orchestrator that independently checks every claim.