Day 4 on the @droid Max plan ($200). @FactoryAI sent a free voucher. I used Lauren Tan's (@poteto) pstack in Cursor for months, and I missed it when I moved to Droid. So I ported it. The part I care about most is the model routing. Each role in pstack gets its own model, and I picked them from Morgan Linton's (@morganlinton) VulcanBench (@VulcanBench ) results instead of going on vibes. Coding goes to GPT-6.1 Sol, which scored 97% of Opus 5.5's best for about a tenth of the cost. Opus handles the judgment calls, and the review panels mix models from three different labs. There's also a budget mode that hands the grunt work to DeepSeek-V4.1-Flash It checks Lauren's repo every week, so her updates keep coming. github.com/NDilanka/pvstack

My iPhone Duo video went viral, so I'm giving away 100 iPhones on X 😛

Day 20 on the @droid Max plan ($200) 14 PRs merged on Acquit today. All Droid, all orchestrated. The day started at PR #115 and ended at #128. Mobile UI got a full pass, the verdict grid actually fits a phone now. Chat cards show PayPal orders and payouts inline. The escrow dashboard has undo receipts with a change log. The big one was disputes. Droid set up a real sandbox buyer dispute through PayPal checkout, funded a job, let the House deliver and get verified, then filed the dispute. Found that PayPal returns a 403 on some dispute reads, so we built a hold path for that. House operators now answer their own buyer disputes with their own PayPal app. Also started on the 3D print verifier. A simulated print farm gets qualified, a judge scores the output, and missing photos fail-close. The validator found three real bugs on the first pass, including one where CPU work blocked the API event loop. Fixed those in round 2. 21 commits across all of it. Ran out of usage at the end but honestly that's a solid day. Still sandbox, still building.

This....

Got selected for the Meta Global AI Developer Hackathon (only 20k spots worldwide). From the kickoff on Oct 26 through the closing ceremony on Nov 9, I’m building Acan in public right here. I’ll share the key steps along the way. Building it with @droid, Meta’s APIs, and the other partner resources they give us. Acan is me rethinking how developers actually interact with AI agents. Chat interfaces got us this far, but I don’t think we should stay stuck with them when we’re developing. There are better ways, and I’m exploring one that feels like it smooths the whole thing out. It’ll be open source once it goes public. Let’s see what happens. Good luck to everyone else building. 🫡

OMG Just got approved for Meta’s AI hackathon. Time to build 🏭

Day 17 on the @droid Max plan ($200) Quick Acquit update for the PayPal AI hackathon. 920 commits in and PR #96 just landed. Today was mostly about making a rejection actually readable. Before, when an agent's fix got rejected you'd only see "hidden tests failed". Now the job page says stuff like "Passed all 48 visible tests. Failed 1 of 6 hidden tests. Escrow still held." And it names which agent made the attempt. The verifier score card is still at 11 of 11 known cheats caught and 20 of 20 honest fixes passed. Right now I'm pointing a free Nemotron model at the live site with a plain "fix this issue" prompt to see if it gets caught overfitting. It only passed 2 of 10 runs earlier so lowkey expecting a rejection 😅 Still all PayPal sandbox. Code freeze is Nov 6. What would you want to see in the demo?

Day 16 on the @droid Max plan ($200) Wanted to know if GLM 5.3 Flash and Nemotron 3.5 Super are actually worth using in Droid, since they're both super cheap in the picker (0.06x and 0.08x). So I benchmarked them. 23 Python tasks from @VulcanBench, 3 runs each, and hidden tests decide if a task counts as solved. Ran DeepSeek V4.1 Flash and GPT-6.1 Sol next to them so I had something to compare against. 690 runs total and my laptop lowkey almost died doing it 😅 Sol is on top at 87%, but it costs 26k to 46k credits a task. The surprise was GLM 5.3 Flash on low. 80.6% solved at about 5.7k credits a task. That's half of what DeepSeek low costs and about a fifth of Sol low. Easily the best value in Droid right now. DeepSeek V4.1 Flash on high is still the sweet spot if you care about getting it right. 86.8% for about 14k credits, so almost Sol level for around half the price, and it's fast too. Nemotron 3.5 Super was the disappointing one. Same score as GLM on high but almost 3x the credits. Honestly couldn't find a job where it wins. So this is how I'm splitting it now. GLM 5.3 Flash low for exploration, mechanical edits and anything high volume. DeepSeek V4.1 Flash high as my default coding worker. Sol only for the hard stuff where getting it right matters more than credits. GLM max barely helped (82.3% for almost 3x the credits of low) so I'm skipping it. It's one Python suite, so take it with a grain of salt.

Day 15 on the @droid Max plan ($200) Today I’m benchmarking GLM-5.3-Flash and Nemotron 3.5 Super inside the Droid harness. My laptop cannot handle the things I’m forcing it to do. I feel bad for you buddy, hang tight. You got this.

Two days of nothing but UI/UX on Acan. Still zero lines of the actual app. Acan draws your system as a factory floor. Services are machines, queues are belts, agents show up on the machine they touch. So @droid fanned out a research swarm. 25 docs, 76,498 words, on agentic UX, supervisory control, attention research, how IDEs juggle workspaces. Claude, GPT and Grok each pitched concepts on their own, then I got interviewed three times to pick what survived. Five prototypes are clickable now. Work orders, clash on the belt, the gate, the bench, and one world canvas holding every project so tab switching can go away. Claude and Grok reviewed each one. Every round came back "not ready", which is the part that actually moved the design. Anyone else this picky about UX before writing code, or am I just avoiding the hard part 😅

Day 15 on the @droid Max plan ($200) Been heads down on Acan, the thing I'm building to actually see what my coding agents are doing. The idea is your system drawn as a little factory. Services are machines, queues are belts, and agents show up on the machine they're changing. Before building anything I had Droid fan out 12 research agents, then had Claude, GPT and Grok each pitch concepts on their own. Turns out "Factorio for agents" is already a whole category (AgentCraft, VibeCraft, Maestri), so that one's out 😅 What survived is the part nobody seems to be doing. You pin real moments from a trace as must-keep and must-not, post that as a work order, and the agents read it back to you before writing any code. If two agents plan to touch the same thing, it shows up as a clash on the belt and you settle it before a single file changes. Three prototypes are clickable now. Work orders, clash on the belt, and one canvas with all my projects on it so I stop tab switching. Light mode too. Two more are being built. Still all fixture data, nothing hooked up to real agents yet.

Day 14 on the @droid Max plan ($200) Started a new project. It's called Acan. The idea is pretty simple. Your running system gets drawn as a little isometric plant. Every service is a machine, traffic moves on belts, and when something strains it lights up. Coding agents like Droid, Codex and Claude Code show up as drones over whatever part they're editing. I'm using Acquit as the first plant since I know every part of it. The PayPal dock, the verifier, the model calls, all of it. It has 4 modes. Live is what's happening right now. Blueprint is where you draw a "ghost" machine that doesn't exist yet and hand it to an agent to build. Simulate runs recorded traffic through a queueing model so you can try 2× traffic or take PayPal down. Replay rewinds the last hour. Replay is my favorite so far. On the demo data PayPal fails first, retries pile up 20 s later, and the API starts failing a minute after. It also checks the deploy from 3 min before and shows it never touched PayPal. One key copies the whole thing as a brief you can paste into any agent. One rule I'm keeping from day one. Agents can propose changes to budgets and checks but they can never edit them. You approve. Why now? Agents made writing code really cheap. Knowing what all of them are doing to your system at the same time is still hard. lowkey think that's the next problem. Right now it's 7 throwaway HTML prototype pages. Trying to nail the UX before the real build. It's local-first btw. Here is a little sneak peek :)

WTH opus, be careful next time😠

Day 14 on the @droid Max plan ($200) Changed how I'm building Acquit for the last stretch of the PayPal hackathon. I'm experimenting with Haiku 5.5 right now. Droid runs as a coordinator now. It writes a short brief for each task, gives it to a Claude Haiku 5.5 worker in its own git worktree, and then GPT-6.1 Sol checks the bigger changes before they merge.Different model family on purpose, so it doesn't just agree with itself. 12 PRs merged so far in this run. Sol sent work back 6 times and honestly everyone was a real catch. My favorite was the job page saying "Nothing was frozen" right above a frozen contract 😅 The other fun one is that the 3D print verifier is already working, about a week ahead of plan. It slices the model with PrusaSlicer, checks infill and scale, and reads signed telemetry from the printer. The printer is simulated for now so I'm calling it that. Also found my laptop's limit. 4 agents at once on 8 GB of RAM and they all died at the same moment. Running 3 now. Still on PayPal sandbox. The code score card catches 11 of 11 known cheats and passes 20 of 20 honest fixes.