Agent Orchestra

Multi-agent fleet conductor for the GitHub Copilot CLI.

Commander sections in harmonyAppend-only ledger motifsSealed score after the run
Agent Orchestra ยท conductor score
$agent conductor premium max on ~/dev/my-repo : final release readiness review
commander-001architecture section
250/250complete
commander-002security section
112 playingactive
commander-003implementation section
5 ledgersin harmony
Shadow Score: sealedFleet Scorecard: play it again with changes

Conduct AI fleets in concert.

0
Commander sectionsvisible panes
0+
Sub-agent ensembleper max fleet
0
Ledger motifsproposal to broadcast
0
Scorecard movementsrepeat decision
The score

Every section gets a part. Every run gets a score.

Agent Orchestra is not a generic swarm. It is a conductor score: visible commander groups, bounded sub-agent fan-out, shared ledgers, live telemetry, and sealed scoring after the work is done.

01
Cue commander sections
state.json + queue

Turn one mission score into five visible Terminal Stampede commander sections with bounded work.

02
Bring in sub-agents
child-agents.jsonl

Each section brings in focused sub-agents while keeping launch proof and telemetry fresh.

03
Play from shared ledgers
collab/*.jsonl

Proposals, reviews, improvements, consensus, and broadcasts become append-only musical motifs.

04
Watch the conductor monitor
orchestrator-commentary.json

Track section status, sub-agent counts, confidence, queue state, and live run health without reading raw logs.

05
Seal the score
scorecard.md

Shadow Score judges quality and Fleet Scorecard answers whether the arrangement is worth running again.

Run score

Hear each section enter the arrangement.

A believable run playback turns the mission, commander activity, ledger motifs, ensemble convergence, and repeat decision into one readable score.

Sample mission score
you

agent conductor premium max on ~/dev/my-repo : final release readiness review

conductor

run-20260505-170318: 5 commander sections cued

section-001

proposal motif: release blockers mapped to auth + billing

section-003

review counterpoint: agrees on auth risk, adds migration caveat

ledger

p5 r18 i11 c8 b7 -> ensemble convergence locked

shadow-score

seal verified - synthesis score: strong

fleet-scorecard

repeat decision: run this arrangement again with changes

proposalpeer_reviewimprovementconsensusbroadcastbundle.jsonscorecard.mdproposalpeer_reviewimprovementconsensusbroadcastbundle.jsonscorecard.md
Conductor monitor

Keep the conductor monitor open while the sections play.

Agent Pulse turns orchestration noise into live orchestration cues: section status, sub-agent counts, ledger tempo, confidence, and recent launches while the run is active.

Commander status3 active / 2 complete
Sub-agents112 running / 480 done
Ledger tempop5 r18 i11 c8 b7
Confidencelive

09:41 commander-002 enters with peer review

09:44 convergence broadcast adopted by 3 sections

09:47 Shadow Score seal still holds pitch

Set list

Use it when one solo agent cannot carry the piece.

Agent Orchestra shines on broad, risky, or cross-cutting work where independent sections find blind spots, compare evidence, and resolve dissonance.

Release readiness

Assign architecture, implementation, tests, docs, and risk to separate sections before a ship/no-ship decision.

Migration risk audit

Let commander sections play compatibility, data flow, rollout, and rollback themes in parallel.

Security posture review

Bring threat modeling, dependency posture, CI, and runtime hardening into the same score.

Architecture pressure-test

Let multiple sections review tradeoffs, surface dissonance, and resolve into the strongest plan.

Large repo onboarding

Map systems, tests, ownership, and sharp edges without listening to raw log noise for hours.

Cross-cutting implementation

Coordinate broad changes where a single part would miss dependencies and failure modes.

Fleet Scorecard

After the work, judge the run.

Fleet Scorecard is the closing score for orchestration: what changed, what rang true, what missed, and whether the same fleet is worth cueing again.

FSS-L4 reference implementation
Four movements. One repeat decision.
Agent Orchestra seals a Fleet Scorecard rubric before commander launch and emits the scorecard during teardown.
01Run score
02Seal the rubric
03Fleet performance
04Evidence index
05Scorecard
What changed?
What converged?
What failed?
Would I run it again?
Repeat verdictRun again with changes
ConfidenceHigh
Failure motifPartial commander bundles
Opening cue

Raise the score with one command.

Install the helper launcher, run the local smoke gate, open Agent Pulse, then cue a mission with a scorecard waiting at teardown.

agent conductor premium max on ~/dev/my-repo : final release readiness review
01Install

Clone, install helpers, and run the smoke gate.

02Open Pulse

Keep the conductor monitor visible while the fleet plays.

03Run mission

Cue commanders and seal the final scorecard.

curl -fsSL https://raw.githubusercontent.com/DUBSOpenHub/agent-orchestra/main/quickstart.sh | bashagent-orchestra-pulse
View repo