Adding tool calls to Parley on a flight, watching Claude make a mess, then drawing the architecture boundary for Claude and watching it execute flawlessly within it. The eureka: Claude Code is the execution harness, Parley is the interface for human to inspect and influence — both running on the same shared filesystem state.
Reflecting on a year of tinkering and adapting to the ever changing world of AI. After a year and 400+ commits on Parley.nvim, I'm convinced the future is humans setting up the environment, with full automation of the coding inside it. Here's my journey so far.
A follow-up parley: log.md is the least controllable file in an agentic issue structure because it depends on the model voluntarily writing during execution. Drop it. Let git history and plan.md diffs serve as the log instead.
A parley conversation on why agentic coding state — issues, plans, logs — wants to live inside the repo, beside the code. Single-threaded topological execution as the sane default; minimal files per issue beat reinventing Jira on disk.
When agentic coding lets a vim user write a custom mouse-aware floating picker because Telescope couldn't do that. The implication for SaaS: programmability stops being a nice-to-have, it becomes the moat.
An iterative spec flow built on a small Makefile and a handful of markdown files — issue, todo, lessons, specs — cuts through the opaque magic of workflow tools. The trick is treating specs as sketches that bound the agent's search space, not exact behavior to confine it. Let AI bringing in world's knowledge!
The week of February 23, 2026: Dorsey's layoffs, Cloudflare's one-engineer Next.js rewrite, Karpathy's tweet. Picking up Conductor, OpenCode, and OpenSpec over a weekend and shipping 15K lines on Parley — agentic coding is no longer the toy it was ten months ago.
LLMs are fundamentally stochastic, but that's not new — the whole ML field always was, and it worked. The interesting question is how to ground stochastic computation in deterministic domains. Two paths: build a deterministic shell around the model (coding), or apply it directly where the domain itself tolerates noise (recruiting, marketing, design).
Mainstream type systems are simultaneously too strong (closed-world assumptions hurt during change) and too weak (can't say "positive integer" or "age between 0 and 200"). For startups in unsettled domains, duck typing wins — types are a luxury you afford after product-market fit.
Andrew Grove treats management as a production system, but the deeper claim is that information workers are managers — their raw material is information, their output is influence. In modern Big Tech, being "a manager" isn't a promotion. It's just different work.
