Ian Jones

Software Engineer
Heavy Music Enjoyer

[Ian Jones @pburghdoom]

@MShteyn5470 Lol, love it!

[Ian Jones @pburghdoom]

Pointed a new domain at Cloud Run yesterday and the DNS leg turned into a perfect snapshot of what infra work is now.

The tool that worked: gddy, a GoDaddy CLI released literally the day before I needed it. My coding agent drove it from our chat session, no dashboard ever opened: dns set swapped the parking A records for Cloud Run's four Google IPs in one shot.

And it already has a day-one bug, a genuinely fun one: dns set implements replace as create-then-delete. DNS itself forbids that for CNAMEs, because two CNAMEs can't share a name. So replacing the www record needed an explicit delete then add. A bug you can only hit on CNAME replace, found by probably its first real user.

The tool you need didn't exist yesterday and has bugs nobody has hit yet. IDK how you plan around that. You just stay light on your feet.

[Ian Jones @pburghdoom]

@MShteyn5470 Yup, logs are both per persona and overall, from the Delphi system itself

[Ian Jones @pburghdoom]

Update on the forecasting tournament (five agent personas, deterministic broker, real prediction-market questions): added a rebuttal wave. Every challenged persona now has to answer every challenge with defend, concede, or amend, and a completeness gate blocks the round until they all do.

Ran it on a Tesla robotaxi market. Eight challenges landed across the five personas; the answers came out 5 amend, 1 concede, 2 defend. The outlier persona (0.43 while the field sat near 0.10) did the thing you'd want from a decent analyst: conceded the part that was actually wrong, defended its method where the attack overreached, rebuilt its reference class, and landed at 0.10.

The honest caveat: an outlier converging to the field is also what folding to peer pressure looks like. The difference is supposed to be in the log (it conceded specific claims, not the room), but whether 0.10 was right is the market's call, not mine.

Consensus ended up far enough from the market price that the paper book (no real money, positions on paper) made its first trades. First time the system has disagreed with a market hard enough to act.

ForecastBench says frontier models have roughly reached superforecaster parity, so raw accuracy is getting boring as a question. The one I care about: can a forecast survive structured attack, and what does it do when it can't defend everything?

[Ian Jones @pburghdoom]

Fun Vue gotcha that cost me a 0.38 CLS: createApp().mount() doesn't hydrate prerendered HTML. It empties the container and rebuilds from scratch. So my prerendered page painted instantly... and then Vue wiped the entire thing and re-rendered it. My "optimization" was shipping a flash of complete content.

The fix needed BOTH halves: createSSRApp (actual hydration) AND mounting inside router.isReady(). Mount before the route resolves and you hydrate an empty router-view against a full page. Mismatch, same wipe.

The humbling part: every code-reading theory about this bug was wrong. Mine was wrong. A fleet of coding agents I pointed at the repo confidently blamed icon fonts, with grep receipts. What settled it was one Puppeteer probe reading the Layout Instability API's shift rects: footer, 305px tall to 0. The whole page was collapsing.

And you can't reproduce it with a warm cache. Local dev always has a warm cache. Field users don't. Real mobile viewport, cold start, watch the rects. Measure the running thing.

[Ian Jones @pburghdoom]

google's agent framework is ADK. the adirondacks are ADK. anyway I now own https://t.co/EeHK3rhzwU, where three scout agents research your climb and a grizzled head guide writes the trail brief. no business model. no roadmap. I am anything if not willing to waste money on stupid shit I build

[Ian Jones @pburghdoom]

Some of this exists but you can't see it from the post: every sweep reports funds mapped, skips with a one-line reason, and the queued frontier, and a grader scores the report against that protocol before it counts.

What genuinely doesn't exist: a stop reason. Sweeps end on budget (2-3 sponsors a run, 8 portcos a profile) and the report never says "stopped on budget, not because the territory was done." The portco cap even drops silently, which is the worst version of what you're describing. Good catch, turning it into an issue.

Handoff owner is just me. Solo shop, the handoff is a chair swivel lol.

[Ian Jones @pburghdoom]

Open sourced BD Desk this week: an end-to-end BD pipeline on Claude's Managed Agents.

The shape: a radar agent walks PE sponsor portfolios every week (one sub-agent per fund, running in parallel), scores every portco against my ICP, and pushes the qualified ones into a small CRM over MCP. Another agent writes the pre-call dossier. A nightly cron drafts the outreach. Every research run gets graded against a rubric and sent back to revise if it comes back thin.

One rule holds everywhere: the system researches and drafts, a human always sends. Pipeline status is never machine-written.

The reason to look even if you don't care about BD: most Managed Agents examples demo one primitive at a time. This composes damn near all of them (graded outcomes, offline evals with pinned baselines, vaults, skills, multi-agent, memory, cron deployments) into one working system. Swap in your own brand pack and it hunts for you.

https://t.co/ic07cvQcs3

[Ian Jones @pburghdoom]

Ran a five-agent forecasting tournament on a real Manifold market this weekend and the contrarian did the one thing contrarians are for: changed a teammate's mind with receipts.

Setup: personas as pure config on Anthropic's Managed Agents. A base-rate agent with zero network egress (can't read news, only reference classes). A news agent with web access but reference-class reasoning banned. The asymmetry is enforced by networking config, not prompts, and it demonstrably worked: 14 evidence claims crossed the wire and no two personas brought the same kind of evidence.

The secret sauce is the challenge gate. A Scientific Reports paper found one confident adversarial agent can drag group accuracy down 10-40%, so my contrarian doesn't get to vibe-disagree. The host-side broker mechanically rejects any challenge that doesn't cite specific evidence already in the shared log.

Whether that beats the market, resolution will tell me. What I know now is the disagreement had to show receipts, and the whole exchange replays from an event-sourced log. Paper trades only. For now lol

[Ian Jones @pburghdoom]

"Why should I let this thing act?" is the real question under every agent platform, and the four big vendors just gave four completely different answers.

Anthropic's bet: isolate the credentials. Managed Agents keeps OAuth tokens in a vault behind a proxy, so the sandbox running generated code architecturally cannot see them.

Vercel bets the code is the danger. Eve treats agent-generated code as untrusted, full stop. It never touches your app runtime; even on my laptop, the Eve leg of my bake-off spins generated code up in its own Docker sandbox. Felt like overkill until I read the published sandbox-escape research on other platforms.

AWS intercepts actions instead: Cedar policy rules at the gateway, checked before every tool call executes. Nothing waits on a human, the law is pre-written.

And Google authenticates identity. Every agent gets a cryptographic ID mapped to authorization policies, plus a registry so only approved agents exist at all.

I've been running the same small agent on three of these (Managed Agents, Eve, ADK) this month. The loops have converged hard; all four ship a managed loop now, that part's commodity. The trust models have not converged at all, and that's the part worth watching, because the runtime you pick quietly picks your threat model too.

[Ian Jones @pburghdoom]

@alexabelonix Thank you!

[Ian Jones @pburghdoom]

open sourced the starter I use for document-grounded AI apps: next.js + claude + pgvector RAG, guardrails on both sides of the model, an eval harness, even a script that generates the stakeholder deck.

two things make it worth a look.

every layer is hand rolled on purpose. no langchain, no wrapper SDKs. the agent loop, the chunking, the grounding check that flags numbers the model can't cite. all of it is small readable code you can rip out and replace. frameworks hide exactly the parts you need to understand.

and the whole domain lives in one folder. swap the pack (prompts, tools, seed data, labels) and the same skeleton goes from credit underwriting to contract review. there are two packs in the repo to prove it.

the live demo runs with zero API keys. no LLM, no embeddings, no DB, still answers with citations. graceful degradation was a design goal and it accidentally became the best demo of the thing.

repo: https://t.co/240pD4PLni
demo: https://t.co/xPfr29P3xA

[Ian Jones @pburghdoom]

Absolutely bonkers. This is amazing.

[Claude @claudeai]

New in Claude Cowork: teach Claude a skill.

Record your screen while you do a task, talk through it as you go, and Claude turns it into a skill it can run again. Find it under Record a skill in the + menu of the Claude desktop app.

Available on Pro, Max, and Team plans.

[Ian Jones @pburghdoom]

Fun fact that stopped me this week: none of the major LLM crawlers execute JavaScript. Not GPTBot, not ClaudeBot, not Perplexity's. Vercel actually tested this across all of them. So a JavaScript SPA can rank fine on Google (Googlebot renders) while being literally blank to every AI system that answers questions about you.

Checked my own app and, yep. https://t.co/KSVCLSGC1Q was a nicely indexed empty shell. curl returned a div and a script tag.

Two evenings to fix: per-route meta and schema at runtime, then build-time prerendering of the 9 public routes with puppeteer, so crawlers get real HTML without executing a single line of JS. (The build sandbox blocked puppeteer's Chromium download, so I pointed it at the Chrome already on the machine. Verified on a throwaway Firebase preview channel before it went anywhere near prod.)

Now curl of any public route returns 21KB of actual content. Same site, same framework, just rendered at build time instead of in the client.

If AI answers are becoming a real discovery surface, "works in the browser" and "exists to a crawler" are different properties. Check yours with curl. Takes ten seconds.

[Ian Jones @pburghdoom]

Building a guitar theory roguelike and this week the map became the actual theory. Slay the Spire style branching act, except the lanes are circle-of-fifths geography: fork left is the flat road, fork right is the sharp road. Navigating the map IS navigating key signatures, whether the player notices or not.

The relics are the part I'm smug about. Roguelike relics bend the rules of the game, and this game's rulebook is literally music theory. So the relics are permission slips: Blues Box makes b3/b7 legal notes, Bebop License unlocks chromatic passing tones. Want to break the diatonic rules? Earn the item that proves you understand them. (Which is just... how jazz works.)

The previous attempt was an autobattler that played fine and taught nothing. Theory as a coat of paint. If the mechanics aren't the curriculum, the fun part routes around the lesson every time.

[Ian Jones @pburghdoom]

https://t.co/X635HIfo0W

[Ian Jones @pburghdoom]

Deleted 3,924 lines from my agent-observability project this week. A full third of core: dispatcher, job tables, coordination locks, the CLI that drove them. Every line had green tests.

The dispatcher existed to claim orphaned sessions. Except sessions created through the real runtime path always carry a run ID, so its match condition could never be true in production. Not rarely. Never. The only caller that ever existed was its own test suite, faithfully verifying behavior nothing could reach.

A study of 35 open-source Java codebases put effectively-dead methods around 16%. For agent-built projects I'd honestly bet higher, because agents bolt on machinery way faster than anyone removes it (adding is the thing they're good at), and every new subsystem shows up pre-wrapped in passing tests that keep it looking alive indefinitely.

CI tells you the code works. It has no opinion on whether the code is reachable. Different question, and the second one is the one that let 4k lines squat in my repo for months.

[Ian Jones @pburghdoom]

my cloud provider capped me at 20 vCPUs (un-raisable, don't ask), the deploy hashes are nondeterministic so every deploy is a full 40-minute redeploy of everything, and my fix was a script that deploys my barely-trafficked side project in polite little batches so the quota doesn't get upset. the future is here and it's a bouncer at an empty club lol

[ClaudeDevs @ClaudeDevs]

Claude Code artifacts can now call MCP connectors, letting you build dashboards and apps that can fetch information and take actions for each viewer on demand.

Available on Pro, Max, Team, and Enterprise plans. Not available on publicly-shared artifacts.

[Ian Jones @pburghdoom]

Swapped the model under Google's ADK for Go from Gemini to Claude. One line changed, whole framework ran. Then an adversarial review of my adapter found three bugs that only exist BETWEEN providers. Google's genai types marshal schema enums UPPERCASE; Anthropic wants lowercase. Multi-agent graphs emit consecutive same-role turns; Anthropic 400s, roles must alternate. And a history can end on a dangling tool_use with no paired result. Also a 400. Gemini tolerates all three shapes, so the framework emits them freely and nothing trips. Until the request hits the stricter provider.

Saw a study that tested structured outputs across 244 models: only 43 passed cleanly on every endpoint. Same bug class. Recursive $refs one API accepts and another rejects, fields required on one door and illegal on the next.

"Model-agnostic" means the interface is shared, not the tolerances. The permissive provider hides your malformed requests; the strict one bounces them. Adapter's in my public repo (adk-tinker) if anyone wants the translation layer, but the keeper is the general rule: your happy path on provider A is a latent bug report on provider B.