Apple Silicon only. Log in if you already have access.
Reconciled 4 ledger exports across 6 workers into a single source of truth. Runway modeled in base, bull, and bear scenarios. Deck draft in progress, holding open roles per your note.
Most agent products run a single loop: one context, one thread, one step at a time. Ultrawork is multi-agent from the ground up — planners delegate, and whatever is independent runs side by side, up to twenty workers at once.
One step at a time, dependent or not — the run takes as long as all of them combined.
Everything queues behind one context window. Long goals stall, drift, or run out of room.
Independent tasks go out together. Anything downstream waits for the handoffs it needs.
A root planner delegates slices to sub-planners and concrete tasks to workers. Independent work goes out at once; dependent work waits for the handoffs it needs.
Most multi-agent systems put one agent in the middle. Every result routes back through its context, so the longer a goal runs, the more that one agent has to hold — and the further it drifts from what you asked for.
The supervisor tracks who has what, and every answer comes back to it. Long goals are exactly where that context fills up.
A planner posts a task to a shared board and moves on. Whoever is free claims it. Nobody tracks anybody, so there is nothing to lose track of — and because the board is a database, the goal outlives the agents working it. This way, we are able to coordinate thousands of agents over months of work.
Two ways to check the claim. A public benchmark against a single-agent harness, and the monthly finance close Homi actually runs on it.
AutomationBench · 204 tasks
Measured against Pi — an open-source single-agent harness — on 204 business-automation tasks, holding the model, the tools and the task prompts constant. One verifier plugin, two passes.
PluginsKimi K3 at max reasoning · two verifier passes · identical tool surface · paired test, p = 0.014 · wins in all six domains
In production at Homi
Ultrawork works the close across dozens of files — building the workbook, reconciling every account, then asking the finance team in Slack when only a person can settle it. It waits for the reply and picks the goal back up, for three days straight.
Signed off against the satisfaction checks and the verified balance sheet numbers.
Codex sessions were given the same work and could not carry it that far.
34 of 34 tasks complete
QUICK JOBS · minutes
LONG-RUNNING GOALS · hours to days
Get a clear view of the metrics that matter. Logs into the observability tools, walks the flamegraphs, and comes back with hotspots plus a short writeup with screenshots.
See more ways to use UltraworkCheckout p95 went from 780ms to 2.1s overnight. Flamegraphs point at two hot spots in the payment path: a retry loop without backoff, and a currency call that stopped hitting cache after Friday's deploy.
Writeup and screenshots are in the doc.
posted in #team-eng-perf with the pack
Ultrawork gets a machine of its own to work on, sits in the channels your team already uses, and takes direction while it's still working.
It can install what it needs, run whatever the job takes, and make a mess doing it. When the work is done the machine is thrown away — nothing it did can reach your files, your network, or anyone else's work.
$ pip install openpyxl
$ python build_workbook.py
board_pack.xlsx · 28.6 KB
Mention it, message it directly, or let it watch a thread. When someone answers, the goal that was waiting on that answer is the one that picks back up. It can wait on you for days without costing anything.
Priya those two are timing differences — book them and close it
Got it. Picking the close back up now.
Tell it something new mid-goal and it re-reads the plan before its next move. Work that no longer makes sense gets dropped, the rest gets re-split. Nothing starts over.
UPDATED PLAN
Model runway with the new hires
Model runway, roles held
Re-run the three scenarios
It isn't a chat that ends when you close the tab. A goal works when there's work to do, parks itself when it needs an answer from you or someone outside, and picks back up the moment that answer arrives — for as long as the job actually takes.
The Q3 close ran 24 days end to end — 62 hours of real work, four pauses waiting on someone to reply.
Coming soon
Gmail, Notion, Calendar, Drive, Jira, Stripe, Snowflake and hundreds more — plus anything with an API, and a machine of its own to run whatever else the job needs.