For teams
Product
Inside Glueprint
Product tour The full workflow, in 16 narrated chapters Workspaces Sessions, code, and terminals together Editor & IDE Edit, preview, and debug beside your agents Tasks & boards Plan the work and hand it to agents Channels People and agents in one conversation Assistants Memory, routines, and ongoing work Git & review Review changes, branches, and GitHub Orchestration Split one job across many agents
Get started
Desktop app Download for macOS, Windows, or Linux Mobile app Your agents on iOS and Android Always-on hosts Set up a VPS or Mac mini with the CLI Supported agents Find the agents you already use
Explore the four steps of AI adoption
Security Docs Pricing FAQ Download free Sign in

Orchestration

One objective.Many agents.

Break a larger job into tasks, hand each one to an agent session on this machine or another, and step in only where a person should decide. Glueprint keeps track of what's ready, what's blocked and what came back.

Download free Watch the orchestration chapter
Sheet OR-001 Plan viewOne run · four tasks
PLAN VIEW · ONE RUN RUN · SHIP LOG SEARCH FOR THURSDAY CHECK ORDER SUBAGENT · DONE EMPTY RESULTS THIS MAC · ATTEMPT 2 RETENTION LINUX BUILD SERVER RELEASE REVIEW COORDINATOR ASKS YOU 01 REPORTS 02 TASKS · WORKERS 03 DEPENDS ON 04 GATE 05 WAITS FOR A PERSON
Coordinators
Sessions · assistants · the CLI
Workers
Subagents · full sessions · other hosts
Decisions
Gates · question cards · inbox
OR-100Anatomy of delegationReal capture · desktop app

Hand off a piece.Watch it come back.

Here the log search session splits off two subagents: one writes the tests while the other checks ordering, and the parent keeps working.

The Build log search session with two subagent cards: Search tests, running and expanded to show its own session timeline, and Order check, completed; above the composer, the plan shows 1 of 4 tasks done with a decision needed, beside Subagents, 2 active.
  1. 01A subagentA child session started from the conversation, with its task, agent, model and status.
  2. 02Its timelineExpand the card to follow the child's prompt, messages and tool calls in place.
  3. 03OpenOpen the child's full chat beside the parent to read it or answer its questions.
  4. 04A finished subagentWhen a child is done, its card says so and the parent picks up its result.
  5. 05The parent keeps goingThe coordinating session keeps working while its subagents run.
  6. 06Plan and subagentsThe run's plan and every child of this session, running or finished, above the composer.
OR-200Runs and tasksReadiness
Session info, Orchestration: the run Ship log search for Thursday's release, 1 of 4 done, with the open Release decision gate and Show question; its four tasks, with Release review waiting on another; and three dispatch attempts with their liveness, one on the Linux build server.

Plan the work.Readiness follows.

A run is one objective broken into tasks, with the dependencies between them. A task is ready when everything it depends on is done and no decision is pending. Glueprint works that out on every read and names exactly what's in the way.

  • Each task shows its state and what's blocking it: another task, or a gate.
  • Each worker shows its attempt and whether it's active, idle, waiting or stalled.
  • A retry is a new attempt. The earlier result stays on record.
  • If a failure leaves nothing able to move, the run says it's stuck, once.
Watch in the tourParallel work and orchestration
Read the guides Orchestration overview MCP tools for orchestration
OR-300Decision gatesHuman review
The Release decision question card in the coordinating session: Search, empty results, and retention are ready. Ship log search in Thursday's release? with the answers Ship Thursday and Hold, and room for your own answer.

Ask a person.At the real decision.

A decision gate holds one task until a person answers. It arrives as an ordinary question card on the coordinating session, and the rest of the run keeps moving while it waits.

  • Offer preset answers, or let the reviewer write their own.
  • Answering resolves the gate, and the task becomes ready once its other dependencies are done.
  • Show question in Session info brings a pending gate back into view.
Watch in the tourParallel work and orchestration
Read the guides Decision gates Permissions and approvals
OR-400Who coordinatesSessions · assistants · CLI

Anyone can run it.Even a script.

The same runs, tasks and gates work whether a session you started is coordinating, an assistant is, or a script on a server is.

  • 01

    A session you started

    Ask the agent to split the work. It starts subagents in its own transcript, or full sessions in their own tabs, and reads what comes back.

    gp_spawn_subagent · gp_create_session · gp_orch_*

  • 02

    An assistant

    An assistant coordinates the same way, within the delegation rules its role allows. What lands in its inbox can wake it.

    gp_spawn_subagent · gp_start_session

  • 03

    A script, with no session

    Drive runs, tasks, workers, gates and the inbox from the glueprint command, with JSON output and fixed exit codes.

    glueprint run create --objective … --auto-close
Watch in the tourParallel work and orchestration
Read the guides Orchestration from the CLI Assistants
OR-500Workers and recoveryEngine guarantees

Workers anywhere.Recovery built in.

The run lives on the host that owns it, workers can run elsewhere, and the engine repairs what a crash or an outage leaves behind.

  • Inline subagents

    A child session in the parent's transcript, on the same host, with its own agent, questions and permissions. Starting one needs no approval.

  • Workers on other hosts

    A full session can run on another machine you've connected. The run stays home; the worker's reports travel back over the relay.

  • Reports that arrive

    A worker's report is saved before it's sent and retried until the coordinator's host has it, so a relay outage only delays it.

  • Stalls noticed

    A worker with no output, idle without reporting, or stuck launching for 10 minutes is flagged, and the coordinator is told.

  • Crash recovery

    At startup and every 30 seconds, lost workers are settled, abandoned runs are closed and missed notifications are sent again.

  • Private tasks

    A run's tasks never touch your boards. A run holds up to 200 of them and is kept for 30 days after it closes.

Workers on other hosts need a plan with cloud relay.

OR-600 Split the work

Give the big job a plan.Keep the last word.

Download Glueprint, start a session on something too big for one agent, and ask it to split the work.

Download free Watch the orchestration chapter