Skip to main content
Back to Projects

AI-Native Workspace

Symphony Workspace

Autonomous Issue Orchestration

An AI-native orchestration workspace that turns tracker issues into isolated implementation runs. Symphony lets teams manage work at the queue level while agents produce code, proof-of-work, and safe review handoffs.

Autonomous runs
Proof-of-work
Linear to PR handoff

Linear

Tracker entry point

Per-Issue

Workspace isolation

Live

Operator observability

Run Preview

Linear issue to safe handoff

Stage 1

Queue Intake

Linear issues are claimed on a bounded cadence, not by ad hoc chasing.

Stage 2

Scoped Workspace

Each issue gets its own runtime, files, hooks, and state contract.

Stage 3

Agent Execution

An LLM-powered agent runs inside the workspace until it reaches a defined handoff state.

Stage 4

Reviewable Handoff

Proof-of-work packages make review visible, fast, and grounded.

Visible

Issue state, retry state, and agent state stay visible together

Visible

Proof-of-work artifacts travel with the run instead of being reconstructed later

Visible

Human review happens at the policy edge, not every turn of execution

Visible

Operators manage queue health instead of babysitting individual sessions

Design thesis

Move project management closer to runtime reality.

Symphony replaces status theater with observable execution: queue state, workspace state, run state, and review state all stay in one operational frame.

01

Problem

Most engineering overhead is still coordination: planning, status updates, handoffs, and reviews exist because humans have to translate intent across roles. Agent tools can write code, but teams still need a reliable way to dispatch issues, isolate execution, track progress, collect evidence, and land changes without supervising every turn.

Coordination Tax

PRDs, planning, status updates, and handoffs consume too much of the engineering loop before code ever ships.

Ticket Tools Stop Too Early

Traditional trackers know the issue state, but not the real execution state of the agent, workspace, run, and evidence trail.

Supervision Does Not Scale

If every coding agent still needs a human sitting next to it, the bottleneck only moves instead of disappearing.

02

Orchestration Flow

The core product bet is that when execution can be delegated to agents, the workspace should stop optimizing for handoffs and start optimizing for observable runs.

Legacy loop

Coordination-heavy delivery

Step 1

Write PRD

Step 2

Sprint planning

Step 3

Status sync

Step 4

Human handoff

Step 5

Review chase

Replace coordination artifacts
with operational state

Symphony

Execution-native workspace

Step 1

Poll Linear

Step 2

Create issue workspace

Step 3

Launch model run

Step 4

Collect proof of work

Step 5

Route to Human Review

Step 6

Land safely

What changes

The visible unit of work becomes the run itself, not the meeting cadence around it.

Why it matters

Reviewers inherit evidence and state, so judgment scales better than supervision.

Operations dashboard preview

Inspired by the real observability surfaces in Symphony: running sessions, retry queue, runtime, and token usage.

Running

8

active issue sessions

Retrying

2

backoff queue

Tokens

148K

in/out across runs

Runtime

6h 12m

aggregate active time

IssueStateRuntimeLast update
SYM-143Human Review42m · 7 turnsValidation green, walkthrough uploaded
SYM-146In Progress17m · 3 turnsReviewing failing CI and retry schedule
SYM-151Merging8m · 1 turnApproved by human, land workflow active

03

Workspace Model

Built Symphony as a long-running orchestration service. It polls Linear for eligible work, creates a workspace per issue, launches an LLM-powered agent runner inside that workspace, feeds a repository-owned WORKFLOW.md contract, and keeps the run moving until it reaches a workflow-defined handoff such as Human Review or Merging. Observability surfaces active sessions, retry pressure, runtime, and token usage so operators can manage the queue with confidence.

Tracker Intake

Symphony continuously reads eligible Linear issues on a bounded cadence and decides what can run now.

Per-Issue Workspace

Every issue gets an isolated working directory so runs stay scoped and deterministic.

Workflow Contract

A repository-owned WORKFLOW.md defines runtime settings, prompts, states, hooks, and quality gates.

Agent Runner

The model runner executes inside the workspace until the issue reaches a defined handoff state.

Observability Surface

Operators see active sessions, retries, runtime, token usage, and issue-level status without inspecting each agent manually.

04

Proof of Work

What the agent hands back

Symphony is designed to make runs reviewable, not mysterious.

CI status and validation output

PR review sweep and reviewer feedback

Complexity analysis for the changed work

Walkthrough videos for human reviewers

Explicit Human Review and Merging handoffs

What remains distinctly human

The value shifts upward: less translation work, more judgment and system design.

Vision: deciding what work matters and why

Architecture: shaping the system so agents can execute safely

Care: defining the quality bar, review bar, and handoff bar

Systems design: reducing coordination overhead by moving intent closer to execution

05

Outcomes

Outcome 1

Turned issue execution into repeatable autonomous runs instead of one-off agent sessions

Outcome 2

Kept execution policy in-repo through a versioned WORKFLOW.md contract

Outcome 3

Produced proof-of-work artifacts including CI status, PR feedback, complexity analysis, and walkthroughs

Outcome 4

Used per-issue workspaces and bounded concurrency to operate multiple runs safely

Outcome 5

Gave operators a live dashboard for active sessions, retries, runtime, and token usage

06

Tech Stack

Elixir/OTP
Phoenix LiveView
Linear GraphQL
LLM Runtime
Isolated Workspaces
Workflow Contracts
Structured Observability
Git Automation