Synergy: A Durable Workspace for Agentic Work That Outlives a Conversation

What Synergy Actually Solves

Most AI agent tools treat a task as a single conversation. When the chat ends, the context, files, and tool state vanish. Synergy is built for the opposite assumption: agent work often outlives one chat session. The README states that Synergy treats agent work as durable workspace state, keeping sessions, agents, files, Browser, tools, and automation connected in one runtime. The target user is an engineer or knowledge worker who runs multi-step tasks that span hours, involve multiple specialist agents, or need to resume after an interruption. The project is built by the Holos team at the Shanghai Innovation Institute, and it is open source under the MIT license. The core value is persistence: a task can move between Web, Desktop, CLI, background execution, and specialist agents without losing its project, history, files, or operating context. That is a specific problem, and Synergy names it directly in its tagline: persistent, recoverable AI agent work.

The Durable Workspace Model

The architecture centers on a local runtime that exposes the same sessions and state across every surface. The README lists the surfaces: Web workbench, Desktop app, CLI, server API, and SDK. A task is attached to an explicit Scope, which is a home or project. Sessions are recoverable, and complete history is kept even when older model context is compacted. That means Synergy does not rely on the model's context window for long-term memory. Instead, it stores session history as durable workspace state. The runtime also keeps project files and a session-owned Browser page in context, so an agent can work across files and a live web page without moving to a separate tool. This is a meaningful architectural choice: the workspace is the source of truth, not the model's conversation. The README also describes a Library for reusable memory and evaluated experiences, with reward-signal analytics per behavioral dimension. That suggests a feedback loop where agents learn from past work, though the README does not detail how the reward signals are computed or applied.

Agent Coordination: From Light Loop to Boss Mode

Synergy offers several coordination patterns, each with a distinct mechanism. Light Loop is for focused, short work that does not need a full plan. Blueprints are durable plans that agents write and can replay. BlueprintLoops are independently reviewed versions of those plans, meaning a separate review step exists before execution. Boss Mode orchestrates a tree of persistent specialist workers. This is a hierarchy of subagents that can persist across tasks. The README also mentions Agenda for scheduling recurring and one-off agent runs on Day/Week/Month calendar views, with run history. The practical implication is that Synergy is not just a single-agent runner; it is a multi-agent orchestration platform. The trade-off is complexity: managing a tree of persistent workers requires careful design of task decomposition and review. The README does not provide examples of how to define a Blueprint or configure Boss Mode, so a new user will need to consult the docs.

Getting It Running: Commands and Config

The Quick Start section gives concrete steps. For the CLI and Web, you install with a curl script: `curl -fsSL https://raw.githubusercontent.com/SII-Holos/synergy/main/install | bash`. Then you configure a model provider with `synergy config wizard`, start the background runtime with `synergy start`, and open the Web client with `synergy web`. A single task can be sent from the terminal with `synergy send "summarize this repository"`. Useful runtime commands include `synergy status`, `synergy logs`, `synergy doctor`, and `synergy stop`. The CLI installer places runtime assets under `~/.synergy/`, and you can change that root with the `SYNERGY_HOME` environment variable. The Desktop app is a separate installer: `.pkg` for macOS, NSIS `.exe` for Windows, and `.deb` for Linux. Portable artifacts exist but do not configure a system CLI. Windows Desktop and CLI releases currently support x64 only. The README warns about channel conflicts: you can have multiple installations, but only one should be the `synergy` command your shell runs. `synergy doctor` detects all installation channels and exits nonzero on conflicts. Upgrades use `synergy upgrade`, and if multiple channels exist, you must specify `--method` with one of `npm`, `yarn`, `pnpm`, `bun`, `desktop`, or `standalone`.

Benchmark Claims and Their Caveats

The README reports a benchmark on DeepSWE v1.1, which is a set of 113 real repository engineering tasks. The claim is that running deepseek-v4-flash with the synergy-max agent lifts Pass@1 from 53% to 67.3% compared to the stock harness, at a cost of $0.54 per task. The README also includes failure anatomy: 76 of 113 tasks fully passed, and 24 of the 37 unsolved tasks missed by only 1 to 2 tests. These numbers are specific and plausible, but they come with a significant caveat. The README states that all costs are computed at API prices in effect before 2026-08-17, prior to DeepSeek's across-the-board price increase and peak/off-peak pricing. That means the cost figure of $0.54 per task is stale. The performance gain might still hold, but the cost advantage could be smaller now. Also, the benchmark compares Synergy's agent to the official mini-swe-agent run of the same model, which is a fair baseline, but it is one model and one benchmark. The README does not provide results for other models or tasks outside repository engineering.

Limitations and Wrong-Tool Cases

The most obvious limitation is the youth of the project. The repository's default branch is `dev`, and the latest release at the time of writing is v3.0.21, which suggests active development but also potential instability. The README mentions plugins, MCP servers, and Channels as extension points, but it does not describe the maturity or count of available plugins. If you need a rich ecosystem of pre-built integrations, Synergy may not have it yet. Another limitation is the channel conflict issue. The README explicitly warns that multiple installations can conflict, and `synergy doctor` exits nonzero when it detects a problem. That is a real operational friction. Also, the Windows support is x64 only, which excludes ARM-based Windows devices. Finally, Synergy is a local-first runtime. If you need cloud-hosted execution or a managed service, this is not it. The README mentions Synergy Link for remote execution, but only in the context of connecting a Holos agent, which adds account identity and messaging. Without a Holos account, remote execution is not available.

Alternatives and the Difference in Approach

The most direct alternative is the stock harness used in the DeepSWE benchmark, which the README calls the official mini-swe-agent run. That is a minimal agent harness that runs a model against a task without durable session state or multi-agent coordination. The difference is fundamental: a stock harness is ephemeral, while Synergy persists sessions, files, and browser state. Another alternative is a general-purpose agent framework like LangGraph or CrewAI, which also support multi-agent orchestration. The difference is that those frameworks are libraries you integrate into your own application, while Synergy is a standalone workspace with a built-in runtime, Web UI, Desktop app, and CLI. If you want to build a custom agent application, a library gives you more control. If you want a ready-to-use workspace that manages sessions and files for you, Synergy is closer to that. The README does not compare itself to these frameworks, so the choice depends on whether you want a product or a toolkit.

Maintenance and Upgrade Cost

The README provides several commands for maintenance. `synergy doctor` checks installation health and channel conflicts. `synergy upgrade` handles upgrades, but it stops and asks for a `--method` when multiple channels are installed. That is a deliberate safety measure, but it means upgrades are not fully automatic in mixed-channel setups. The README also mentions that the curl installer and npm package postinstall warn about other channels and never auto-uninstall them. That is a conservative approach that avoids accidental data loss, but it puts the burden on the user to resolve conflicts. The license is MIT, which gives you broad freedom to modify and redistribute, but it also means there is no warranty or support guarantee. The project is backed by the Shanghai Innovation Institute, which suggests institutional support, but the README does not describe a formal maintenance schedule or long-term roadmap. You should expect to handle configuration and troubleshooting yourself, especially given the `dev` default branch.

Editorial conclusion

Adopt Synergy if you run agentic tasks that must survive across sessions, surfaces, and model context compaction, and if you value local ownership of projects and data. Skip it if you need a mature plugin ecosystem or if you rely on benchmark numbers that predate a price change. Before adopting, verify your model provider's compatibility with the config wizard, test the Blueprint and Boss Mode workflows on your own tasks, and check the current state of the official and local plugin registries.

Official sources

1. Official README
2. Project repository
3. Release notes