The Discussion Board
Session 23 of Living on the Frontier, in person at The Kannas Hotel in Chiang Mai — and a deliberate format change. Last week’s feedback was clear: less news recap, more technical discussion. So this week the agenda came from the room before the room ever sat down — topics collected from the community chat all week, compiled Friday night, voted on live Saturday morning. Pick one, go deep, move on.
Everything below is the room’s agenda, not a manifesto: each item compresses a thread from the community chat, counter-takes included. Every topic is redacted on purpose — no reading ahead. Each one gets decrypted live when the room picks it.
The board
Trusting code you don’t read
01"Less looking at code" — the trust spectrum. Tolerance for not reading AI code varies wildly by industry: fintech moving millions and healthcare teams still human-review every PR, while startups auto-merge. Very few have found the formula for auto-merged AI code in production that stays high-quality long-term. The sharpest framing from the chat: your own systems = you are the user — would you apply the same mechanics in a deployed SaaS with a million paying users? Does your AI care about a massive failure at 3AM?
02Uncle Bob's pivot. His stated strategy is to not read his agents' code — extreme constraints instead of review, unit tests as the cage. And an honest arc from the host: I cited him for years, tuned him out during the early AI wave, and this year I'm listening again.
03Automated docs systems. The GitHub Agentic Workflows case: a team turned merged features into cross-repo docs PRs with scoped permissions and SME review — 82 of 82 merged.
04Comments as liability. Code is a liability; comments are too. Stale comments are precise but inaccurate — and now they mislead your agents, not just your teammates.
Specs and the source of truth
05Spec as the source of truth. The week's biggest thread. The claim: code changes too much to be trusted; the spec has the last say — not code, not comments, not tests. Positions from the chat: plain English is too ambiguous, so make specs verifiable (Gherkin-style, or executable markdown that errors when code drifts); an artifact is only a source of truth if everything downstream reproduces from it with zero manual intervention; structured-English methodologies as the middle ground, with specs updated in the same PR as the code. Against all that: Anthropic engineers have said in public videos that they don't work from specs, while Cloudflare has published how it enforces engineering standards mechanically. And a live query — does dev with Claude models simply look different from dev with open-weight models? Opening poll: who in the room runs an SDD framework today?
Model behavior
06Models that delete versus models that route around. A first, reported in the chat: a frontier model put onto another model's codebase said "we need to delete all of this — this hardening breaks these features." One model deletes; another never deletes a line by design, it routes around and makes it bigger. What that means for pairing them — and members' tuning tricks for running strong models on lower effort settings.
07Sandboxes versus cages. Another sandbox escape during cybersecurity testing made the rounds this week. The chat's provocations, paraphrased: if sand always finds its way out of a sandbox, the honest word is cage — and a classifier softens behavior, it doesn't contain it. Continues Session 22's escape story — hardened VMs versus everything else.
The tools we run
08The harness debates. A week of frank field reports in the chat: members switching CLI harnesses and saying exactly why, bug lists pulled from real session logs, update-caution, and the counter-take that the whole GUI-wrapper category is solving a problem that shouldn't exist. Who's on what, who moved, and why.
09Multiplayer AI rooms. Humans plus their agents in one shared space keeps resonating. The community now has a live experiment running — open platform, members pulling their own agents in.
10Picking a chat platform for the agent era. Self-hosting your data versus a vendor holding it and charging per seat; agents designed at the center versus tacked on; how hard it is to install an agent at all.
11Platform security and the decentralization question. The chat's skeptics versus its builders on the agent-chat experiment: the security trade-offs of relay-routed architectures and keypair identity, what a real security boundary looks like for a community platform, and the comparison one member drew between decentralized identity and passkeys.
The economics
12Commodified intelligence. The chat's framing of a strange week in pricing: a very cheap, very capable model signals a price hike, and coverage shared in the group described a hyperscaler offering a frontier-class model at bargain rates reportedly tied to data-sharing terms. One member's dichotomy: either someone made a real efficiency breakthrough, or the market is carrying a large inference markup.
13The token economics roundtable. Subscriptions versus API budgets, members' router setups that make a $40/month budget go a long way, the efficiency of the open-weight challengers, and the psychology of unlimited tokens.
14Cheap-model clusters versus one expensive reviewer. One member's benchmarks across 100 PRs, shared in the chat: a clustered team of three cheap models hit 94% bug capture at a fraction of the false-positive rate and 7× cheaper than the premium single reviewer.
Undercard — if the room is fast
15Quick hits. Cost-efficiency and budgeting notes from the chat; daily harness-performance trackers; and a member-built in-browser VM builder that shipped this week — maybe we take it for a spin live.