Agent orchestration and workspaces

In one paragraph

Agent orchestration covers the layer above a single agent: the workspace, runtime, and coordination surface that lets multiple agents, or agents and people, do work together. Some tools give you a desktop IDE with per-agent Git worktrees and a diff reviewer. Others run a shared event log so humans and agents post to the same channels. A few are AWS-specific control planes with Terraform stacks and compliance tooling. One adds literature search and experiment tracking for research workflows. What counts here: workspaces, IDEs, and platforms where several agents, or agents and people, run and coordinate work.

The main approaches

Agent IDEs

A desktop application wraps one or more CLI coding agents. Each agent task gets its own Git worktree and branch so parallel work stays isolated. The IDE adds diff review, browser preview, and PR integration on top. superset runs Claude Code, Codex, Gemini CLI, and others this way, in parallel. OpenResearch also wraps coding agents in a desktop app but targets research workflows, adding literature search against alphaXiv, bioRxiv, and PubMed plus a local SQL experiment store.

Shared workspaces

Agents and humans share a common communication and event surface rather than separate tools. buzz runs a Nostr relay where messages, Git patches, CI results, and workflow steps all land in one signed event log. OpenDots lets you give each named agent its own isolated browser and file workspace, and connects the agents to a document workspace with human-in-the-loop approval. loom is a self-hosted single-binary daemon that launches and supervises coding-agent sessions over REST, MCP, and a Vue dashboard, with every route derived from one Rust declaration.

Platforms on AWS

Full infrastructure stacks for running agents inside an AWS account. sample-agentic-platform is a monorepo with five Terraform stacks, Helm charts, and 25+ lab notebooks covering EKS or Bedrock AgentCore compute paths. sample-agentic-value-accelerator targets financial services teams and adds a React/FastAPI control plane, 35 FSI-specific multi-agent POC implementations, and compliance checklists for 14 frameworks.

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["superset-sh/superset"]
  f1 --> r2["alphaXiv/OpenResearch"]
  f2 --> r3["marin-community/loom"]
  f2 --> r4["block/buzz"]
  f2 --> r5["CopilotKit/OpenDots"]
  f3 --> r6["aws-samples/sample-agentic-platform"]
  f3 --> r7["aws-samples/sample-agentic-value-accelerator"]

Where the new ideas are

buzz: uses Nostr keypairs to give every agent and human the same signed identity in one event log, so git patches, CI results, and chat messages are all auditable in the same place.
loom: derives REST routes, MCP tools, and CLI commands from a single Rust declaration per operation, so the server, CLI, and MCP adapter cannot drift out of sync.
OpenResearch: records every experiment run in a local SQL database so agents can read past results before deciding what to try next, closing the loop between hypothesis and evidence.
OpenDots: can give each named agent a persistent isolated browser, file workspace, and shell through a container supervisor, with credentials derived server-side per agent.
sample-agentic-value-accelerator: includes a DCF business-case builder and a maturity assessment alongside agent code, connecting operational tooling to business justification.

Side by side

Repo	Approach	Multi-agent	Human-in-the-loop	MCP support	Infrastructure included
aws-samples/sample-agentic-platform	Platforms on AWS	Yes	Not stated	Yes	Yes, 5 Terraform stacks + Helm charts
superset-sh/superset	Agent IDEs	Yes, parallel worktrees	Yes, diff review before commit	Yes, MCP server included	Not stated
marin-community/loom	Shared workspaces	Yes	Yes, PR and CI review via browser	Yes	Not stated
aws-samples/sample-agentic-value-accelerator	Platforms on AWS	Yes	Not stated	Not stated	Yes, Terraform on ECS Fargate
block/buzz	Shared workspaces	Yes	Not yet, approval gates in progress	Not stated	Yes, Nostr relay, Postgres, Redis, S3
alphaXiv/OpenResearch	Agent IDEs	Not stated	Not stated	Not stated	Not stated
CopilotKit/OpenDots	Shared workspaces	Several named Dots; delegation between them is further work	Yes, human approval step on pages and non-read-only tools	Yes	Not stated

How the idea moved

flowchart LR
  n1["started Apr 2025<br/>aws-samples/sample-agentic-platform"]
  n2["started Oct 2025<br/>superset-sh/superset"]
  n3["started Jan 2026<br/>marin-community/loom"]
  n4["started Feb 2026<br/>aws-samples/sample-agentic-value-accelerator"]
  n5["started Mar 2026<br/>block/buzz"]
  n6["started Jun 2026<br/>alphaXiv/OpenResearch"]
  n7["started Sep 2026<br/>CopilotKit/OpenDots"]
  n1 --> n2 --> n3 --> n4 --> n5 --> n6 --> n7

started Apr 2025 · sample-agentic-platform · Adds: Two concrete AWS compute paths (EKS with Helm, and ECS plus Bedrock AgentCore Runtime) sharing one agent interface and gateway layer, with five Terraform stacks, 25+ lab notebooks, JWT auth, and OpenTelemetry telemetry.
started Oct 2025 · superset · Adds: A desktop IDE that runs any CLI coding agent in a per-task Git worktree with diff review, per-workspace browser preview, Design mode for click-to-edit UI changes, PR line feedback, and scheduled automations.
started Jan 2026 · loom · Adds: A self-hosted single-binary session daemon that derives every REST route, MCP tool, and CLI command from one Rust declaration, supervises coding-agent sessions in worktrees, and tracks PR and CI state.
started Feb 2026 · sample-agentic-value-accelerator · Adds: An FSI-focused AWS control plane with a React/FastAPI UI, 35 multi-agent POC implementations, a DCF business-case builder, an App Factory that generates agent code from plain-language briefs, and compliance checklists for 14 frameworks.
started Mar 2026 · buzz · Adds: A self-hosted Nostr relay that puts agents and humans in the same signed event log, combining channels, Git hosting, YAML workflows, and CI status into one searchable audit trail; most starred in this theme.
started Jun 2026 · OpenResearch · Adds: A research-agent harness that adds literature search against alphaXiv, bioRxiv, and PubMed, records every experiment in a local SQL database for agents to query, and supports compute across local, SSH, Slurm, Kubernetes, and cloud targets.
started Sep 2026 · OpenDots · Adds: A template where each named agent can get a persistent isolated browser, file workspace, and shell through a container supervisor, connected to a document workspace with human-in-the-loop page approval.

Easily confused

This theme is about coordination layers, not about what a single agent does internally. A tool that defines how one agent reasons, uses tools, or calls APIs belongs in agent frameworks or tool use. The concern here is how multiple agents or agents and people share state, review work, and hand off tasks. Similarly, this is not about model training or fine-tuning: every tool runs agents at inference time against existing models.

Gaps nobody has filled

Nothing here measures the latency or cost of coordinating across many concurrent agents at production scale.
Beyond Git worktrees and merges, no approach resolves conflicts when two agents modify the same file at once; OpenDots’ revision checks only stop stale page edits from overwriting newer ones.
Regulatory-framework checklists appear only in the AWS-based financial services platform. The self-hosted tools offer audit trails but no mapping to regulatory controls.
Agent identity is handled inside each platform (Nostr keypairs in Buzz, JWTs validated against an IDP cert in the AWS sample), and no shared identity scheme spans platforms.
Memory services exist within individual platforms (such as aws-samples/sample-agentic-platform’s memory gateway), but no approach provides a shared long-term memory store that agents from independent workspaces or different frameworks can read and write with consistent semantics.

RationalStuff v3