In one paragraph
Several agents, or agents and people, working on one job need a place to run, a way to stay out of each other’s files, and a way to talk. That comes in different forms: desktop IDEs that give each task a Git worktree, shared workspaces where agents and people use the same channels or repository, delegation layers that fan work out to agent CLIs, and AWS reference platforms for production. What counts here: Workspaces, IDEs and platforms where several agents, or agents and people, run and coordinate work.
The main approaches
Agent IDEs
Desktop apps that wrap the CLI agents you already use. Each task gets its own workspace, terminals and review tools, and you stay in the loop.
- superset runs Claude Code, Codex, Gemini CLI, Cursor Agent or any terminal agent in a Git worktree per workspace, with a diff viewer, browser panes and an MCP server.
- Paradev is the same app under its newer name (formerly Superset). Its README still uses the superset name for the CLI binary and release files.
- c11 is a macOS terminal multiplexer. Agents control the layout themselves over a socket.
- OpenResearch is a desktop harness for research. It adds literature search and a local experiment store to a coding agent.
Shared workspaces
Places where agents and people share one record of the work, such as channels, a repo or documents, instead of one person watching many terminals. Some are servers or skills that start, watch and talk to coding agents.
- buzz puts people and agents on one Nostr relay with channels, git events and YAML workflows.
- OpenDots is a template for persistent agents that can each have their own computer, a document workspace and approval steps.
- Agensh has agents claim sub-tasks and merge work through Gitea and Mattermost, with no central orchestrator.
- loom is a single-binary server that launches sessions in worktrees and exposes REST, MCP, a CLI and a dashboard.
- AutoSwarm installs as a skill in your current agent and hands subtasks to other installed agent CLIs.
- overmind turns production agent traces into datasets, evals and fine-tuned models.
Platforms on AWS
Reference architectures for running agents in your own AWS account.
- sample-agentic-platform gives Terraform stacks and Helm charts for EKS, or ECS with Bedrock AgentCore.
- sample-agentic-value-accelerator is a control plane for financial services teams, covering planning through governance.
Map of the theme
flowchart LR t["Agent orchestration and workspaces"] t --> f1["Agent IDEs"] t --> f2["Shared workspaces"] t --> f3["Platforms on AWS"] f1 --> r1["ParadevAI/Paradev"] f1 --> r2["superset-sh/superset"] f1 --> r3["Stage-11-Agentics/c11"] f1 --> r4["alphaXiv/OpenResearch"] f2 --> r5["marin-community/loom"] f2 --> r6["block/buzz"] f2 --> r7["overmind-core/overmind"] f2 --> r8["HKUDS/AutoSwarm"] f2 --> r9["microsoft/Agensh"] f2 --> r10["CopilotKit/OpenDots"] f3 --> r11["aws-samples/sample-agentic-platform"] f3 --> r12["aws-samples/sample-agentic-value-accelerator"]
Where the new ideas are
- Agents driving the workspace. Superset has an MCP server so agents create and manage workspaces. c11 exposes every panel over a socket so agents split areas and spawn sub-agents.
- No central orchestrator. Agensh has workers claim sub-tasks and merge progress asynchronously. Its README reports test-pass rates rising as agent count grows on five ProgramBench tasks.
- One log for people and agents. Buzz keeps messages, patches, CI results and workflow steps in one signed, searchable event log.
- One declaration per operation. Loom derives its REST routes, MCP tools, CLI commands and authority boundaries from a single Rust declaration, so they cannot disagree.
- Orchestration as a skill. AutoSwarm lives inside the agent you already run instead of being a new app.
- Research grounding. OpenResearch lets agents read past experiments in a local SQL database before choosing the next one.
- Closing the loop. Overmind carries traces through curation, fine-tuning and serving in one stack.
Side by side
| Repo | Approach | Form | Agents it works with | Language |
|---|---|---|---|---|
| aws-samples/sample-agentic-platform | Platforms on AWS | Terraform stacks, Helm charts, EKS or ECS with Bedrock AgentCore | Five sample agents built with Strands or LangGraph | Python |
| ParadevAI/Paradev | Agent IDEs | Desktop app, plus CLI, SDK and MCP server | Claude Code, Codex CLI, OpenCode, Cursor Agent, Gemini CLI, Amp Code, any CLI agent | TypeScript |
| superset-sh/superset | Agent IDEs | Desktop IDE with Git worktree per workspace | Claude Code, Codex, Gemini CLI, Cursor Agent, any terminal agent | TypeScript |
| marin-community/loom | Shared workspaces | Single-binary server with REST, MCP, CLI and Vue dashboard | Coding agents launched from a goal string | Rust |
| aws-samples/sample-agentic-value-accelerator | Platforms on AWS | React and FastAPI control plane on ECS Fargate | LangGraph and Strands Agents SDK POC agents | TypeScript |
| block/buzz | Shared workspaces | Nostr relay, Tauri desktop app, buzz-cli | Agents calling buzz-cli as tool calls | Rust |
| overmind-core/overmind | Shared workspaces | Hosted or self-hosted platform with SDK | Agents instrumented over OpenTelemetry | Python |
| Stage-11-Agentics/c11 | Agent IDEs | macOS terminal multiplexer | Agents controlling panels over a socket | Swift |
| alphaXiv/OpenResearch | Agent IDEs | Desktop app (macOS, Windows beta, Linux) | Claude Code, Codex, OpenCode, Cursor, Google Antigravity | Rust |
| HKUDS/AutoSwarm | Shared workspaces | Skill and CLI | Codex, Claude Code, OpenCode, Cursor, GitHub Copilot CLI and others | Python |
| microsoft/Agensh | Shared workspaces | Harness on Gitea and Mattermost | Copilot harness (github/copilot-cli) in the authors’ experiments | Python |
| CopilotKit/OpenDots | Shared workspaces | TypeScript template with document workspace | Named specialist Dots with MCP server permissions | TypeScript |
How the idea moved
flowchart LR n1["started Apr 2025<br/>aws-samples/sample-agentic-platform"] n2["started Oct 2025<br/>ParadevAI/Paradev"] n3["started Oct 2025<br/>superset-sh/superset"] n4["started Jan 2026<br/>marin-community/loom"] n5["started Feb 2026<br/>aws-samples/sample-agentic-value-accelerator"] n6["started Mar 2026<br/>block/buzz"] n7["started Mar 2026<br/>overmind-core/overmind"] n8["started Apr 2026<br/>Stage-11-Agentics/c11"] n9["started Jun 2026<br/>alphaXiv/OpenResearch"] n10["started Jul 2026<br/>HKUDS/AutoSwarm"] n11["started Aug 2026<br/>microsoft/Agensh"] n12["started Sep 2026<br/>CopilotKit/OpenDots"] n1 --> n2 --> n3 --> n4 --> n5 --> n6 --> n7 --> n8 --> n9 --> n10 --> n11 --> n12
- started Apr 2025 · sample-agentic-platform · Adds: two compute paths, EKS with Helm or ECS with Bedrock AgentCore Runtime, that share one agent interface and gateway layer, so teams can compare them on working code.
- started Oct 2025 · Paradev · Adds: a Git worktree, terminals, diff viewer and browser preview per task around whichever CLI agent you already pay for, with a CLI, SDK and MCP server so agents can drive it too.
- started Oct 2025 · superset · Adds: a desktop IDE with per-worktree browser previews, a diff viewer, scheduled automations and an MCP server that lets agents create and manage their own workspaces.
- started Jan 2026 · loom · Adds: a self-hosted session supervisor where every REST route, MCP tool, CLI command and authority boundary comes from one Rust declaration per operation.
- started Feb 2026 · sample-agentic-value-accelerator · Adds: a control plane for financial services teams that spans business-case planning, agent building, security, operation and compliance checklists for 14 frameworks.
- started Mar 2026 · buzz · Adds: a shared Nostr event log with keypair identity, so messages, patches, CI results and workflow steps from people and agents sit in one searchable audit trail.
- started Mar 2026 · overmind · Adds: a loop from live trace collection through dataset curation, fine-tuning and OpenAI-compatible inference serving in one platform.
- started Apr 2026 · c11 · Adds: a terminal multiplexer where every panel is scriptable over a socket, so agents split areas, spawn sub-agents, drive a browser, message each other and post status to a sidebar.
- started Jun 2026 · OpenResearch · Adds: literature grounding against alphaXiv, bioRxiv and PubMed plus a local SQL store of every experiment, so agents read past results before choosing what to try.
- started Jul 2026 · AutoSwarm · Adds: orchestration as a skill inside the agent you already use, which selects, runs in parallel and verifies work across locally installed agent CLIs.
- started Aug 2026 · Agensh · Adds: decentralized coordination where agents gather context, claim a sub-task, act, verify and merge asynchronously through Gitea, Mattermost and a shared context mechanism.
- started Sep 2026 · OpenDots · Adds: persistent agents that can each have an isolated browser and file workspace, wired into a document workspace with a human-in-the-loop approval step.
Easily confused
- A single coding agent. These projects do not replace Claude Code or Codex. Superset, Paradev, Loom and AutoSwarm wrap or direct agents that already exist. AutoSwarm even runs inside one.
- An agent framework. LangGraph or Strands define how one agent reasons. The AWS samples use them as parts, while workspaces such as Superset, Buzz and Loom are about where agents run and how they coordinate.
- Two separate products. Paradev and Superset are the same app, renamed. The Paradev README still uses the superset name for its CLI binary and release files.
- Observability alone. Overmind scores traces, but it also curates datasets and fine-tunes models. It sits next to the workspaces, not inside one.
Gaps nobody has filled
- No head-to-head comparison of worktree-per-task IDEs with shared-repo or decentralized coordination on the same task.
- Scaling evidence for many agents working together comes from one benchmark set, Agensh’s ProgramBench runs. There are no scaling measurements for worktree IDEs, shared channels or skill-based delegation as agent count grows.
- No single view across a desktop IDE, a server-side session supervisor and a shared workspace for one team’s mixed agents.
- No worked cost comparison between the EKS path and the AgentCore path, even though aws-samples/sample-agentic-platform sets them side by side.