The right work, on the right seat and model.
Work arrives understood: a brief with the why on top, sized before anyone starts, and routed on what matters: who knows that part of the code, who has the headroom and the time, and which model does it well for the least. Engineers get a list and a wall that is an exit. Leads get the flow of work, summed by area.
sleds uninstall is one line out. Hooks, status line, token, skill, Codex provider: every trace. The board keeps the estimates and actuals.The best way through your list.
Most engineers never touch credits and rarely hit a limit. Every engineer has a list and finite hours. sleds works out what is ready and what is blocked, what to run now and what can run in the background while you think, which model each step needs, what can run overnight, and which teammate owns the thing you are about to start. Your window and budget are constraints it respects, not the point of the plan.
| when | request | size | note |
|---|---|---|---|
| now | SSO redirect loop on the Acme tenant | M | worth high · $4–11 · Fable |
| now | Review PR #409 from last night | S | signed · 340/340 tests |
| background | Migrate lint rules to the new config | S | worth low · $1–3 · Haiku |
| overnight | Backfill tests for billing webhooks | M | worth medium · $9–24 · ½ price |
| teammate | Rework the export API | L | Dana owns billing · headroom |
| held | Rename the internal enum | S | worth unstated · $3–8 |
A better list, and never leaving the terminal.
Routing isn't only for the lead. For the engineer it means the right work shows up ready, the model is picked per step, the wall is an exit, and anything that can wait runs overnight at half price. Your laptop, your install, your numbers.
For you
Three things were put in front of Jake overnight: one sent by a teammate, one proposed by the rules because he landed the last three rows in that area, one review GitHub asked for. He takes the first from the Slack DM with one word.
for you · 3 · oldest first sent Help and Support · $4–14 proposed flag fallback · you landed the last 3 review #1341 · +412 −38 › take
Sized, with a model per step
Before he starts he sees the band, the cap, and which model fits each step. The plan runs on Sonnet; the one tricky migration step steps up to Fable.
size M · $4–14 · 9–29 min · 78% plan sonnet 5 · 4 steps step 3 fable 5.1 · schema migration cap $20 · nothing bills past it
Background while he thinks
The lint migration fits Haiku and runs in the background while he reviews #1341. He never leaves the terminal.
background lint rules → new config · haiku est $1–3 · 6–14 min ✓ landed · $1.40 · 9 min
A wall that is an exit
His weekly bar hits 100%. The session is snapshotted and six exits appear with prices. He hands it to Dana, who owns billing and has headroom; the snapshot goes with it.
Wall: usage limit reached. Session snapshotted.
5 Hand off to Dana (idle 42m) · Slack + /sled pickup
› /sled 5
Handed off to Dana. The session resumes there.Overnight at half price
What can wait until morning is queued for the batch tier, capped, with the musher answering what it can from the spec.
overnight 2 jobs · batch tier · ½ price window 23:00 → 06:30 cap $12 each
Signed, in the morning
Both jobs delivered, verified and signed. The requesters hear in their own threads. The next estimate is a little better.
06:20 delivered · PR #409 · $6 actual · est $9–24
signed sonnet-5 · 340/340 tests
08:00 digest → #eng · 6 delivered · 0 at a wallOnly what was put in front of you.
Sent by a person, proposed by the rules, or asked for in GitHub, each with the one action it needs. Take it from the board or reply take in Slack. Pass it on or say not me without a meeting; the routing learns from both. Every routed hand-off comes with its reason; decline without explaining, and the sender is told either way.
At 90% the seat tells you. At 100% you have six exits, priced.
Before you start you see the time, the tokens and where the run leaves your window for each model. The status line shows the reset, not just a percent. At 90% the seat says so, inside your session.
sled ▮▮▮▮▮▮▮▮▮░ 90% · resets 1h 12m · credits $14 this ENG-482 · M · $4–11 · 20–55 min · cap $20 fits Fable (band) · Sonnet +40% time ⅕ cost next ENG-479 fits overnight · Dana has headroom for ENG-471
14:02 ENG-482 · wall · seat jake@ at 100% 14:02 policy continue-under-threshold · worst case $11 < $20 · allowance ok 14:02 continued on the platform key · chosen by policy, not by a person 14:36 landed · $7 actual · receipt posted
Wall: usage limit reached. Session snapshotted. Provider: Weekly limit reached. Resets Thursday 9:00 am. Remaining work: $4.10–9.30 · 20–55 min · 82% 1 Continue now platform key · allowance left $38.00 2 Hold until reset free · resumes 9:00 am EDT · saves ≈ $6.70 vs continuing now 3 Run overnight batch tier · ½ price 4 Switch model sonnet 5 · platform key · $0.90–2.10 · same time · rest of session 5 Hand off to Dana (idle 42m) · Slack + /sled pickup 6 Defer to reset runs itself 9:00 am EDT · same model · full price · $4.10–9.30 Reply /sled 1, /sled 2, /sled 3, /sled 4, /sled 5 or /sled 6. wall jd7k2m9qa41p > /sled 1 Continuing on the platform key. Worth high · policy allows · $4.10–9.30 against $38.00. Logged.
Agent work you can read, and a way through every wall.
What the agent did, which model ran it, what it cost, how that compares with the estimate, and a signature. The record for when you have to defend the work.
Hand off, hold, switch model, run overnight, or continue within an allowance. The snapshot goes with the work.
Time, tokens and window position for each model, before you pick one.
At the wall the exit follows the org's policy, and the audit log says so.
The receipt verifies model, timing and outputs. It protects you in a compliance review too.
Accept, decline or bounce a hand-off with a reason. Quiet hours. DMs, not channel noise.
Which model you finished work in, where context churned, when you hit walls. Yours alone.
"Here is a task on your list that fits an overnight run; you haven't tried one." Unused capacity pointed at work.
sleds uninstall removes every trace. The board keeps the estimates and actuals.
Not the spread of spend. The flow of work.
How much work is being requested, how much is in flight on agents versus people, how much is landing with a receipt, where the pile-up is, and what it all cost. The same screens your engineers use, one zoom out, summed by area.
23 rows · 14 on agents · 9 on people
31 PRs · every one with a receipt
4 in review · 2 at a wall
78% of sizes landed in band
Illustrative · a team of fifty. Worth attaches to work, never to a person.
Doing is cheap now. Review is the queue.
In the backlog we studied, an agent's plan ran in minutes and the PR sat open for hours: a two-to-four-times gap that is people, not compute. The pile-up shows where that queue is, by area, with the worth waiting in it, so a lead can move a reviewer or pair two people instead of buying more tokens. No dollar claimed here until it is measured.
| area | what is happening | worth | cost |
|---|---|---|---|
| billing | 4 waiting on review · 2 days median | $48k waiting | — |
| exports | 2 at a wall · policy: hold | $26k · resets thu 9:00 | — |
| auth | 6 in flight · 4 on agents | $31k in flight | $140 so far |
| platform | 9 landed this week · all receipted | $35k landed | $180 |
| unvalued | 5 rows with no worth stated | ask the filer | — |
Set four numbers once. Nobody asks you for a limit again.
The workspace's monthly cap on the platform key. The number finance already wanted and never had a place to put.
Per seat, per month: what a seat may continue on the key past its own bars. Raised on request, in the audit log.
Auto-continue under this. A worst case under the threshold and inside the allowance continues without a question. Worth high can raise it.
Hold, unless worth is high. What happens at 100% when nobody answers. The audit log says the policy chose.
The page is for glancing. Your AI is for asking.
The same aggregates the overview shows are what the connector answers with, by area, whichever AI is asking.
sled: what's stuck in billing4 rows waiting on review, 2 days median, $48k worth waiting. Move a reviewer or pair two people.sled: what landed this week and what did it cost31 PRs, $212k worth, $1,140 cost, 78% of sizes in band.sled: which rows have no worth5. Filed by three people. Here is the list, with "ask the filer" on each.The week, in the channel.
What landed, what it was worth and what it cost, what is waiting and where, what hit a wall. Leads read it where the team already talks, by team and area.
This week on the sled illustrative landed 47 rows · $212k worth · $1,140 cost in band 78% of sizes · calibration moving up overnight 19 rows at half price · $71 saved walls 12 hit · 12 priced · 0 waited review billing is the queue · 2 days median unvalued 5 rows · ask the filer
Nothing changes in your session until you ask for an estimate or hit a limit.
sleds init finds your handles (git email, GitHub login, agent account) and asks you to accept them, backs up your agent settings once, connects the agent to your seat, and installs the status line, the /sled skill and the hooks. sleds doctor shows what is wired, including your cockpit. sleds off hands the agent back to your seat directly.
The CLI, hooks, status line and proxy client are open source. "Your laptop, your install" is easier to believe when you can read the code.
curl -fsSL sleds.ai/install | sh$ curl -fsSL sleds.ai/install | sh ✓ sleds 0.9 · mac arm64 $ sleds init git email dana@acme.dev ↵ accept github dana-acme ↵ accept agent claude code · dana@ ↵ accept cockpit cmux detected ✓ status line · /sled skill · hooks · settings backed up once $ sleds doctor seat ok · cockpit cmux · key workspace
Start with one engineer.
Useful the same day, and it costs the workspace nothing. Request an invite for the team.
curl -fsSL sleds.ai/install | sh