These aren’t single-prompt problems. They need multiple specialized agents coordinating in sequence, in parallel, and sometimes in cycles.<\/p>\n\n\n\n
Most frameworks approach this by making the orchestrator itself an LLM\u2014an agent that dynamically plans which agents to call, in what order, and with what inputs. That works when the task is exploratory. But for workflows with known structure (and in practice, many of the most useful workflows do have known structure), dynamic orchestration adds cost, latency, and unpredictability that can work against you.<\/p>\n\n\n\n
Conductor<\/strong><\/a> is an open-source CLI (MIT license, Microsoft org) that takes a different approach: you define your multi-agent workflows in YAML, and the routing between agents is deterministic. Jinja2 templates and expression evaluation handle conditions and branching. The orchestration layer consumes zero tokens. The structure is fixed at definition time\u2014and that’s the point.<\/p>\n\n\n\n We kept building multi-agent workflows\u2014code review pipelines, design document generation, research assistants\u2014and writing the same glue code every time: Python scripts stitching prompt chains, ad hoc retries, manual state between steps, no good way to version-control the workflow itself.<\/p>\n\n\n\n We looked at other tools, such as Microsoft Agent Framework (MAF), Microsoft’s primary SDK for building agents in code, which covers many of the same primitives. Conductor<\/a> is a different surface for similar patterns: a YAML-first CLI for teams who want to compose agents and tools without writing SDK code. Declared, diffable, and as readable as a CI\/CD pipeline.<\/p>\n\n\n\n We also wanted to separate concerns that keep getting mashed together in multi-agent systems:<\/p>\n\n\n\n Conductor is the result<\/a>: YAML workflows, isolated agents, and a routing graph you can see before anything runs.<\/p>\n\n\n\n Every Conductor workflow is a YAML file that declares agents, their prompts, models, inputs, outputs, and routing logic. Workflows are version-controlled, diffable, and reviewable, the same way you’d treat infrastructure-as-code or CI\/CD pipelines.<\/p>\n\n\n\n Routing between agents uses Jinja2 templates and expression evaluation. First matching condition wins. A workflow can loop hundreds of times through an evaluator-optimizer cycle without the routing layer consuming any tokens. This is what separates Conductor from dynamic orchestration: the workflow topology is declared, not discovered at runtime.<\/p>\n\n\n\n Conductor<\/a> supports GitHub Copilot and Anthropic Claude as providers, with per-agent model overrides. You can mix them in a single workflow: run claude-haiku-4.5 for classification, gpt-5.2 for research with MCP tool access, and claude-opus-4.6-1m for complex reasoning. Each agent gets its own session with no shared conversation state.<\/p>\n\n\n\n Static parallel groups run multiple agents concurrently with configurable failure modes (fail_fast, continue_on_error, all_or_nothing). Dynamic for each groups process variable-length arrays in parallel with batched concurrency. Results are aggregated and available to downstream agents through template expressions.<\/p>\n\n\n\n Not every step needs an LLM. Script steps run shell commands directly, capturing stdout, stderr, and exit codes into the workflow context. A code review workflow can run pytest between the “implement” and “review” steps. Routes can branch on exit codes. No model invocation, no token cost.<\/p>\n\n\n\n Human gate steps pause execution, present options in a Rich terminal UI or the web dashboard, and route based on the response. Approval workflows, review checkpoints, interactive decision points: they’re part of the workflow graph, defined the same way as any other step.<\/p>\n\n\n\n Conductor includes a web dashboard that visualizes execution in real time. An interactive DAG shows the workflow topology with animated edges for execution flow. Each node is clickable, showing the agent’s prompt, model, token usage, cost, activity stream, and output. Human gates work directly in the browser. Background mode (–web-bg) starts the dashboard, prints the URL, and returns control to the terminal.<\/p>\n\n\n\n Three context modes control what each agent sees: accumulate (all prior outputs), last_only (just the previous step), and explicit (only named dependencies). The default is accumulated, but for larger workflows, explicit mode cuts token consumption significantly. Being deliberate about what each agent sees turned out to matter more than we expected.<\/p>\n\n\n\n Plugins follow the Agent Skills<\/a> open standard, bundling reusable skills and MCP server configurations that agents can use. Reference them from Git repos or local paths. Workflow registries let teams share and version workflows: configure a registry once, then run workflows by short name.<\/p>\n\n\n\n Max iteration limits and wall-clock timeouts prevent runaway execution. Dry-run mode previews the execution plan without calling any models. Conductor<\/a> doesn’t replace your editor, CI system, or agent framework. It’s a CLI that reads YAML, calls models, and produces structured output. It plugs into what you already have:<\/p>\n\n\n\n The most common pushback is “what about dynamic orchestration?” Fair question. If your task needs to restructure itself based on what it discovers, let the LLM decide what comes next. But the workflows we keep reaching for (review loops, research pipelines, plan-then-implement) have known structure. We’d rather have predictability, cost control, and auditability than replanning flexibility. Conditional routing and loop-back patterns cover more ground than you’d expect.<\/p>\n\n\n\n Each agent gets its own session, system prompt, model, provider, and temperature. No shared conversation bleeding. This seems like overhead until you’re debugging a workflow where step 4 is mysteriously influenced by step 2’s output. Explicit context flow makes multi-agent systems tractable.<\/p>\n\n\n\n The engine uses a pub\/sub event system for all output. The terminal renderer, web dashboard, and any future consumers subscribe independently. More work upfront than printing to stdout, but it decoupled the execution engine from the presentation layer in a way that keeps paying off. Adding the web dashboard required zero changes to the workflow engine.<\/p>\n\n\n\n We considered Python APIs, JSON schemas, and visual builders. YAML hit the sweet spot: readable, structured, diffable in pull requests, and familiar to anyone who’s written a GitHub Actions workflow or a Kubernetes manifest.<\/p>\n\n\n\n MIT-licensed, developed in the open from day one.<\/p>\n\n\n\n Contributions welcome: provider integrations, workflow examples, plugins, docs, bug reports.<\/p>\n\n\n\n Install:<\/strong><\/p>\n\n\n\n # macOS \/ Linux<\/em> Run your first workflow:<\/strong><\/p>\n\n\n\n conductor run workflow.yaml –input question=”What is Python?”<\/p>\n\n\n\n Visualize it:<\/strong><\/p>\n\n\n\n conductor run workflow.yaml –web –input topic=”AI in healthcare”<\/p>\n\n\n\n Conductor requires Python 3.12+ and works with GitHub Copilot or Anthropic Claude. The repository<\/a> has documentation, example workflows, and a getting-started guide.<\/p>\n\n\n\n Multi-agent workflows are becoming infrastructure: repeatable, versioned, shared across teams. We chose deterministic orchestration because for the workflows we build most often, known structure is the whole point.<\/p>\n\n\n If you’re stitching together agent pipelines with glue code, give Conductor a look.<\/p>\n\t\t\t\t\t<\/div>\n\n\t\t\t\t\t\t\t\t\t\t\tThe problem with multi-agent workflows today<\/h2>\n\n\n\n
\n
Key capabilities of Conductor<\/h2>\n\n\n\n
YAML-defined workflows<\/h3>\n\n\n\n
workflow:\u00a0
\u00a0 name: design-review\u00a0
\u00a0\u00a0entry_point: architect\u00a0
\u00a0
agents:\u00a0
\u00a0 - name: architect\u00a0
\u00a0\u00a0\u00a0 model: claude-opus-4.6-1m\u00a0
\u00a0\u00a0\u00a0 prompt: |\u00a0
\u00a0\u00a0\u00a0\u00a0\u00a0 Create a design document for: {{\u00a0workflow.input.purpose\u00a0}}\u00a0
\u00a0\u00a0\u00a0 output:\u00a0
\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0file_path: { type: string }\u00a0
\u00a0\u00a0\u00a0 routes:\u00a0
\u00a0\u00a0\u00a0\u00a0\u00a0 - to: reviewer\u00a0
\u00a0
\u00a0 - name: reviewer\u00a0
\u00a0\u00a0\u00a0 model: claude-opus-4.7\u00a0
\u00a0\u00a0\u00a0 prompt: |\u00a0
\u00a0\u00a0\u00a0\u00a0\u00a0 Review the design at {{\u00a0architect.output.file_path\u00a0}}\u00a0
\u00a0\u00a0\u00a0 output:\u00a0
\u00a0\u00a0\u00a0\u00a0\u00a0 score: { type: number }\u00a0
\u00a0\u00a0\u00a0\u00a0\u00a0 approved: { type:\u00a0boolean\u00a0}\u00a0
\u00a0\u00a0\u00a0 routes:\u00a0
\u00a0\u00a0\u00a0\u00a0\u00a0 - to: $end\u00a0
\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0\u00a0 when: \"{{\u00a0output.approved\u00a0}}\"\u00a0
\u00a0\u00a0\u00a0\u00a0\u00a0 - to: architect\u00a0<\/code><\/p>\n\n\n\nDeterministic routing, zero token overhead<\/h2>\n\n\n\n
Mix providers and models per agent<\/h2>\n\n\n\n
Parallel execution<\/h3>\n\n\n\n
parallel:
- name: researchers
agents: [academic, web, technical]
failure_mode: continue_on_error
routes:
-to:synthesizer <\/code><\/p>\n\n\n\nScript steps<\/h3>\n\n\n\n
Human gates<\/h3>\n\n\n\n
Web dashboard<\/h3>\n\n\n\n
Context control<\/h3>\n\n\n\n
Plugins and workflow registries<\/h3>\n\n\n\n
Safety limits<\/h3>\n\n\n\n
conductor validate<\/code> catches schema errors, missing references, and unreachable agents before anything runs.<\/p>\n\n\n\nWorks with your existing tools<\/h2>\n\n\n\n
\n
What we learned<\/h2>\n\n\n\n
1. Determinism is a feature<\/h3>\n\n\n\n
2. Agent isolation pays for itself<\/h3>\n\n\n\n
3. Events over logs<\/h3>\n\n\n\n
4. YAML is the right level of abstraction<\/h2>\n\n\n\n
Open source and ready to use<\/h2>\n\n\n\n
\n
How to start using Conductor today<\/h2>\n\n\n\n
curl -sSfL https:\/\/aka.ms\/conductor\/install.sh | sh
# Windows (PowerShell)<\/em>
irm https:\/\/aka.ms\/conductor\/install.ps1 | iex<\/p>\n\n\n\n![\"Close](\"https:\/\/opensource.microsoft.com\/blog\/wp-content\/uploads\/2026\/03\/CLO25b-Developer-WorkAnywhere-3-1024x768.jpg\")
\t\t\t\t<\/div>\n\t\t\t\n\t\t\tStart using Conductor<\/h2>\n\n\t\t\t\t\t