Factory on X

Praise is concrete. People describe tagging Factory under a video in a Slack thread and getting 15–20 shorts out with copy written, prototyping a marketplace in twenty minutes, or building a database for finding founders; they name prebuilt sandboxes, 20-plus built-in skills and built-in image generation as what clicked. One user says their tokens stretch further than on plans they already pay for; another cites the choice of models and generous usage limits.

The mission strand is its own thing: set a job up once and it "runs until it's done" — still cranking after a laptop closed, a whole product from one prompt.

The undecided say it "looks cool" and they should try it; a few bought the cheapest plan or are on a trial and call mission mode unique. The sceptics argue a greenfield one-shot is the easy win and the real test is an existing repo with its own opinions, judging a claim on the rerun rather than the first build. Those who tried it and left give two reasons: one says it is not their vibe and will not renew; the other moved off because they could not sign in with subscriptions they already pay for.

Usage is the recurring sore point. People on Max describe being 79% through the month in 12 days, or out of the monthly allowance in two weeks; they say cheap models are fine but on a frontier model "it's over", that overnight runs ate the month, and that cloud computer hours stop mid-work and make migration painful.

Quota mechanics are the second: an automatic switch to Droid Core on Kimi K3 once the five-hour Standard limit hit, expensive defaults such as GPT 5.6 Sol and Opus 4.8, a bloated custom skillset chewing usage, and a Droid Core reset date later than the renewal date. Cost complaints say cheap workers iterate until the bill matches a frontier model, that tasks cost more than token rates suggested, and that the monthly price is thin.

CLI and SDK reports: the CLI not pulling missions that show in the desktop and web apps, an SDK release behind the CLI's permission requests that broke a third-party harness, a session id that changed overnight, and a TUI taking 30 seconds to open with copying and mouse support still mangled. Support complaints cite unanswered emails and no follow-up.

Claude Code is the most named reference: one user says Factory is still "the GOAT" for local work and has finally caught up on cloud, and urges people to look outside Claude Code and Codex; another calls Droid the best one-shot harness they have used; a third dismisses the field as wrappers around Claude Code whose value is build over buy.

The T3 Code strand is about lock-in: people would rather standardise on a harness-agnostic setup than be loyal to any one harness, and one built an RPC-based Droid adapter in their T3 Code fork and stopped caring about the TUI.

Devin prompts a straight question — is it better than Cognition's? — and a $20 Droid is weighed as the agent that keeps going after the limit. Subscription stacking is arithmetic: $200 Factory beside $200 OpenAI, $300 Grok and a cancelled $200 Cursor, or advice to take Claude at $100 plus Droid at $100. One person paying for Grok will not renew Factory; another came back to Factory from Grok. Cursor is contrasted as steer-heavy against missions, Codex is named as another tool to try, and one switched from Antigravity because Google held back model access.

Model choice is where the talk is. People trade settings — Opus 5.5 on medium orchestrating, auto worker, Sol 6.1 on high for verification — argue Sonnet 5.5 holds up as a worker on long runs, ask what auto picks and what it considers, and whether open-weight workers wander enough to cost the same as frontier ones. The efficiency advice is about routing: "fewer unnecessary calls" and clear instructions matter more than short prompts, and one asks a model to slim a skill down.

Cost measurement is its own argument: one person logged a mission at 11 sessions and 21.7M tokens, 79% workers and 93.5% cache, and asks others to measure theirs; others want cost per task once retries are counted.

Orchestration questions ask what handoff between agents looks like and whether running 16 to 32 agents in one project holds up. Context questions ask how memory is shared across chats, answered that same computer means shared context, and whether a delegated task keeps its session when you reconnect from a phone. Others ask free or paid, what $20 buys, and whether Haiku 5.5 or the newest DeepSeek checkpoint is in.

Limits are the ask with the largest group behind it: don't cut usage when moving to MAX, drop the 5-hour limit, and loosen the weekly and monthly allowances and cloud computer hours, which one person says stop mid-work at 80 hours a month.

Linking work back is next. People like starting from a Jira ticket because it removes a handoff, but want the return trip: the run, the proposed diff and the test result attached to the same issue so a reviewer can see what "done" means without chasing tools. One wants Droid to pick up only tickets tagged for agents, with a clear definition of done; another wants every delegated task to carry an id it must answer on, because silence looked like progress.

The rest are single requests: running other people's browsers inside Factory Environments, full OIDC support so integrations don't depend on long-lived static credentials, one Droid to manage all the others, a finished mission that leaves something production-ready rather than a demo, and a Factory–LGND integration so Factory's bots can see LGND's data.