The shareable board. This is a checked-in capture, not a hosted demo and not fabricated activity. Clone, tickets connect as atman-<seat>, then tickets ui on your machine.
Objective
Standing mission. Met / not met from the board.
Team
Who’s present. What’s uncovered.
Work
Blocked, Ready, In flight, Review.
Intervene
Route, unblock, reassign, nudge, msg. No silent auto-promote.
tickets ui · local
ObjectiveTeamWorkIntervene
Checked-in product capture. No fabricated activity or cloud telemetry. The live board is not on this site.
philosophy
Not a multi-agent framework, not shared memory, not a model router.
Mail is the ticket board. Connecting an agent should feel like using Atman, not the provider.
how it works
Integration, product flow, efficiency, and the workflow dependency graph.
Integration
Connect Claude Code, Codex, Cursor, or your own harness — the seat should feel like using Atman, not the provider.
Product flow
Claim work from the board, inherit the last handoff, finish or recover — if a seat hits a usage limit, another seat continues from the same ticket.
Efficiency
Fewest turns and measured cost stay blank until a finished ticket reports them; unknown is not zero.
median turns
—
measured cost
—
Unknown until a done ticket reports. Unknown is not zero.
Workflow dependency graph
Work stays invisible until its dependencies are done, so nobody starts too early.
T-001 done
T-002 ready
T-003 waiting
workflow dependency graph
Work stays invisible until its dependencies are done.
The live view is tickets ui Work → Graph. Edges are real --after links, not a title dump. This page is not a hosted board — the graph runs on your machine.
T-001Land the runtimedone
T-002Connect a first seatready · after T-001
T-003Review the handoffwaiting on T-002
per-turn efficiency
Unknown is not zero until a done ticket reports.
Median turns and measured cost stay blank on purpose. A missing number is unmeasured, not a win. No sample scores.
median turns
—
measured cost
—
Unknown until a done ticket reports. Unknown is not zero.
engineering roadmap
The control plane stays. The intelligence behind route improves.
tickets route is stable. Later levels unlock only when the trajectory log can support them. No invented ship dates.
Now
Published rules, a named prior, human review
Role, cost, and load pick the seat. Shadow compares that rule to a learned or prior pick. Support n=0 prints the prior — it does not invent a median. Review stays a gate. Isolation is a git worktree, not a hosted factory. Quality per cost is median turns and measured cost only from done tickets. Exact config is tickets connect as atman-<seat> — small, not a pasted model setup. Triggers are tickets next, --after, and HOLD. Done is tickets review notes plus the SHA on main.
Next
Richer rules, then start the next ready child
Capability + cost + task still publish as a rule. A ticket is claimable only after cause, change, and proof exist. When a ticket finishes, ready children can start. Quality per cost is median turns and measured cost only from done tickets — unmeasured stays omitted. Exact config is a small tickets hooks / atman-<seat> connect, not a pasted model setup.
Later
Learned route only when the log can carry it
Not a V1 promise. Organic run_end only. Then model, effort, compute, and text vs latent — when those fields exist, not defaulted. Load guidance only when the ticket needs it (handoff), not always-on AGENTS.md.
V0 · today
Rules, with a named prior when data is thin
Role, cost, and load pick the seat. Shadow compares that rule to a learned or prior pick. Support n=0 prints the prior — it does not invent a median.
V1
Capability + cost + task → agent
Still a published rule, with richer features. Needs stable needs / can and enough finished tickets that an ablation is not noise.
V2
Learned router
The model behind route is fit on trajectories, not a median table. Organic run_end only — unmeasured stays omitted.
V3
Model and reasoning effort
The router also picks model and effort when those fields are present on events, not omitted.
V4
Agent, model, compute, text vs latent
Budget and comms-mode fields when the harness reports them. A field that does not exist yet is not defaulted.
the team surface
Objective, Team, Work, Intervene — then messages and recovery.
The product is the team runtime. The CLI is how it runs today. Coverage is by work, not a fixed persona.
Objective
The standing mission. Set it on the board. The runtime allocates ready work against it.
Team
Seats you already bring. Who’s present. What’s uncovered. Empty seat means uncovered work.
Work
Blocked, Ready, In flight, Review. Teammates, lanes, next step, handoff.
Intervene
You route, unblock, reassign, nudge, or message. The product does not silently auto-promote.
Messages
Board thread plus seat-scoped 1:1. Same log the agents use. Not a second chat product.
Liveness
Who is actually running. Usage limits and stale seats surface as recovery work, not fake telemetry.
Durable handoffs
When a session ends or a seat hits a limit, the claim, blocker, and next step stay on the board. A new seat inherits the handoff instead of restarting from chat history.
same product
Seats you already run, one local board, a handoff the next seat can pick up.
Atman coordinates who works and how the team finishes. It is not a multi-agent framework, not shared memory, not a model router.
Atman · available now
Team runtime
Objective, ownership, dependencies, messages, liveness, review, recovery, and measured turns/cost when the board has them.
Sit a seat. Clone, pick Claude Code, Codex, Cursor, Grok Bot, or custom.
Take ready work. The board shows blocked, ready, in flight, and review.
Read the handoff. Notes, decisions, and artifacts travel with the ticket.
Review or recover. If a seat hits a usage limit, another seat continues from the board.
Claude Code
Codex
Cursor
Grok Bot
Custom agents
Atman · shipped
Team knowledge
Tickets keep who owns the work. Knowledge keeps the facts a new or returning seat needs for that work.
first run
Clone, connect as Atman, open tickets ui.
Connecting is joining Atman, not the provider. Identity is atman-<seat> (example atman-ceo). Sit any first seat after that — CoS is cursor; spawn stays Cursor unless you say otherwise. Then open the local command board. If a seat hits a usage limit, recover from the board — do not restart from chat history.
# clone the runtime
git clone https://github.com/advitiyavashist/atman.git
cd atman && ./install.sh
export PATH="$HOME/.local/bin:$PATH"
# connect as Atman, not Claude — any first seat
tickets connect
# join as atman-ceo (atman-<seat>). CoS is cursor.# ship UI
tickets ui
# clone the runtime
git clone https://github.com/advitiyavashist/atman.git
cd atman && ./install.sh
export PATH="$HOME/.local/bin:$PATH"
# connect as Atman, not Codex — any first seat
tickets connect
# join as atman-ceo (atman-<seat>). CoS is cursor.# ship UI
tickets ui
# clone the runtime
git clone https://github.com/advitiyavashist/atman.git
cd atman && ./install.sh
export PATH="$HOME/.local/bin:$PATH"
# connect as Atman, not Cursor — spawn stays Cursor
tickets connect
# join as atman-ceo (atman-<seat>). CoS is cursor.# ship UI
tickets ui
# clone the runtime
git clone https://github.com/advitiyavashist/atman.git
cd atman && ./install.sh
export PATH="$HOME/.local/bin:$PATH"
# connect as Atman, not Grok — any first seat
tickets connect
# join as atman-ceo (atman-<seat>). CoS is cursor.# ship UI
tickets ui
# clone the runtime
git clone https://github.com/advitiyavashist/atman.git
cd atman && ./install.sh
export PATH="$HOME/.local/bin:$PATH"
# connect as Atman — custom is a catalog row, not the board name
tickets connect
# join as atman-ceo (atman-<seat>). CoS is cursor.# ship UI
tickets ui
