Microsoft Agent Framework for Go
Microsoft's open-source framework for building production-grade AI agents and multi-agent workflows in Go, with orchestration, observability, and Foundry integration.
- Source repo
- microsoft/agent-framework-go
- Stars
- ★ 695
- Last updated
- today
- License
- MIT
- Primary language
- Go
- FA score
- 82/100 · Good
At a glance
- How it runs
- Works with
- Universal · cross-platformOpenAI API
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- Go teams building production multi-agent orchestration beyond a single prompt: sequential, concurrent, group collaboration, and custom routing
- Not a fit if
- Teams needing declarative agents, DevUI, AF Labs, or handoff orchestration, which are not yet implemented in the Go SDK
- Development teams not using Go; .NET and Python implementations live in the upstream repository
- Teams unwilling to accept public-preview status for production systems
- Website
- aka.ms/agent-framework ↗ set on GitHub
- Maintenance
- Commits in the last 30 dayslast commit 2026-11-22
- Source review
- 82/100 · Good
What does this agent do, and when should you use it?
Microsoft Agent Framework for Go (microsoft/agent-framework-go) is the Go implementation of Microsoft Agent Framework, in public preview and evolving outside the upstream core alongside the .NET and Python implementations. It ships agent, provider, workflow, and middleware packages for building, orchestrating, and deploying AI agents and multi-agent workflows. The framework integrates with a broad ecosystem including Microsoft Foundry, Azure OpenAI, OpenAI, MCP, A2A, AG-UI, and the GitHub Copilot SDK. Workflows are graph-based, supporting sequential, concurrent, and group-collaboration patterns, conditional routing, subworkflows, checkpointing, streaming, human-in-the-loop, and time-travel. The Go SDK covers core agents, tools, middleware, workflows, observability, and interoperability, while declarative agents, DevUI, AF Labs, and handoff orchestration are not yet implemented.
It calls Foundry and OpenAI models through provider packages such as provider/foundryprovider and provider/openaiprovider, running agents via a.RunText(ctx, ...). The workflow package builds graph-based multi-agent orchestration with sequential, concurrent, and group-collaboration patterns, conditional routing, subworkflows, checkpointing, streaming, and human-in-the-loop control. agent/middleware.go provides request/response processing, OpenTelemetry, tool approval, and automatic tool calling. agent/skills, agent/textsearch, and agent/memory supply domain skill bases, automatic RAG retrieval with citations, and vector-backed chat memories. The vectordata package offers vector stores including Azure AI Search and an in-memory connector with hybrid search, and the embedding package generates provider-independent embeddings. Observability flows through provider/otelprovider and workflow/observability/opentelemetry for distributed tracing. Examples run from examples/01-get-started through examples/03-workflows and examples/05-end-to-end.
- Go teams building production multi-agent orchestration beyond a single prompt: sequential, concurrent, group collaboration, and custom routing
- Platform teams needing checkpointable, restartable, observable (OpenTelemetry) workflows with human-in-the-loop control
- Enterprise developers building project-backed Foundry agents or invoking existing server-side Foundry agents
- Knowledge applications that need automatic retrieval and citation of external knowledge (RAG) or session-backed vector memories
- Integrations requiring MCP servers, A2A, or AG-UI interoperability
- Teams needing vector search, hybrid search, and embedding generation infrastructure
Key setup facts
Pulled from the install and usage notes below. The project README has the current values.
- Environment variables
FOUNDRY_PROJECT_ENDPOINTFOUNDRY_MODEL
How do you install or deploy this agent?
Install with go get:
go get github.com/microsoft/agent-framework-goBefore running samples, authenticate to Azure/Foundry (az login is recommended for development) and set the relevant environment variables.
How do you use this agent?
A minimal Foundry agent that writes a haiku:
package main
import (
"cmp"
"context"
"fmt"
"os"
"github.com/Azure/azure-sdk-for-go/sdk/azidentity"
"github.com/microsoft/agent-framework-go/provider/foundryprovider"
)
func main() {
endpoint := os.Getenv("FOUNDRY_PROJECT_ENDPOINT")
model := cmp.Or(os.Getenv("FOUNDRY_MODEL"), "gpt-4o-mini")
token, err := azidentity.NewDefaultAzureCredential(nil)
if err != nil {
panic(err)
}
a := foundryprovider.NewAgent(endpoint, token, foundryprovider.ModelDeployment(model),
foundryprovider.AgentConfig{
Instructions: "You are a helpful assistant.",
},
)
ctx := context.Background()
fmt.Println(a.RunText(ctx, "Write a haiku about the Microsoft Agent Framework").Collect())
}Set FOUNDRY_PROJECT_ENDPOINT (optionally FOUNDRY_MODEL) and authenticate with az login first. See examples/01-get-started, examples/03-workflows, and docs/agent-harness.md for more.
What are this agent's strengths and limitations?
- Officially maintained by Microsoft with a broad ecosystem: Foundry, Azure OpenAI, OpenAI, MCP, A2A, AG-UI, and GitHub Copilot SDK, reducing lock-in to one provider
- Graph-based workflows include checkpointing, restartability, streaming, human-in-the-loop, and time-travel patterns aimed at production
- OpenTelemetry integration at both the agent/provider and workflow layers for distributed tracing and debugging
- Complete retrieval and memory stack: agent/skills, agent/textsearch (RAG), agent/memory, vectordata, and embedding packages
- Public preview: declarative agents, DevUI, AF Labs, handoff orchestration, and Foundry-hosted deployment are not implemented in the Go SDK, which trails .NET
- Using third-party servers, agents, or non-Azure direct models is at your own risk, with usage, costs, and compliance responsibility on you
- Responsible AI mitigations (metaprompting, content filters) must be implemented by you in production
- By default a feature-usage bitmask is appended only on approved Azure/Foundry HTTPS destinations; disable with AGENT_FRAMEWORK_FEATURE_MASK_DISABLED if unwanted
How does this agent compare with similar options?
The upstream microsoft/agent-framework repository hosts the .NET and Python implementations with broader product integrations and more complete features; the Go SDK is a public preview that currently evolves outside the upstream core, with closer alignment expected. A feature comparison is documented at docs/dotnet-go-sdk-feature-comparison.md.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Microsoft Agent Framework for Go This agent | 82 · Good | Library / SDKFree + model costs | ★ 695 | today | Go | OpenAI API |
| Microsoft Agent Framework Hands-On Samples | 54 · Major gaps | Library / SDKFree + model costs | ★ 397 | 5mo ago | Jupyter Notebook | — |
| Microsoft Agent Framework | 70 · Some gaps | Library / SDKFree + model costs | ★ 15k | 5d ago | Python | OpenAI API |
| AgentOS | 64 · Some gaps | Library / SDKFree + model costs | ★ 691 | 1d ago | TypeScript | Claude Code · OpenAI API · Claude API |
How does FollowAgents rate this agent?
Why each dimension lost points
Framework itself ships no dangerous defaults: the telemetry doc states the feature mask is emitted only to allowlisted Azure hostnames, never third parties, with env-var opt-outs; the README assigns third-party risk explicitly to users and recommends managed identity for production. Least privilege is reasonable but no code-level permission audit is present in the sample (deducted). Tool-approval middleware is claimed, indicating a confirmation mechanism, but its source is not in the packet, so 2. Data-flow transparency is excellent: per-bit mask semantics, emission conditions, hostname allowlist, opt-outs. Sensitive data: docs claim the mask contains no user data and users are told to audit data flows, but no masking/encryption code evidence. Dependency security: pinned, well-known modules in go.mod, but no vulnerability-scan configuration shown. External effects require explicit env vars/credentials; no destructive defaults. Rollback: only workflow checkpointing is mentioned; no framework-level rollback documented — clear deduction. Attribution is clear: MIT, Microsoft copyright, MSRC reporting path.
Self-consistency is strong: README, telemetry doc, and verifyexamples behavior align, unimplemented features are explicitly flagged, no contradictions. Dependency availability: pinned reputable modules, but AG-UI uses a pseudo-version commit dependency (unstable provenance, deducted). Failure messaging: catalog and verify tools produce structured, explicit errors; README has an auth troubleshooting table; runtime error quality is not evidenced in the sample, so 2.
Audience/scenarios well covered: a dedicated fit-check section plus production, orchestration, RAG, and memory scenarios. Capability boundaries are a highlight: unimplemented Go SDK features (handoff, declarative agents, AF Labs, DevUI, Foundry hosting) are enumerated alongside a .NET-Go parity doc. Trigger precision: the sampled SKILL.md files have clear descriptions and argument hints, but they are repo-internal tooling skills rather than the product's user-facing trigger surface, so 2. Environment fit is good: quickstart, env-var table, az login guidance, and a verifier tool target the Go/Azure environment.
Information architecture is excellent: complete TOC, progressive example categories, links to concrete paths. Install notes are one usable line but lack version/release-channel guidance (deducted). Naming is consistent (provider/, agent/, workflow/), but public-preview status means API instability, so 2. Examples and FAQ are rich: progressive samples, troubleshooting tables, transparency FAQ. Known limitations are honestly listed. License is complete (MIT, Microsoft copyright). Versioning/changelog: no CHANGELOG and no version policy despite a v0.1.0 User-Agent — clear deduction. Maintenance responsibility is clear: CONTRIBUTING, SECURITY, SUPPORT policies and an MSRC vulnerability path.
Output usability is good: deterministic example verification and buffered structured JSON catalog output. Marginal value is high: no equivalent Go counterpart to MAF .NET/Python exists; the symbol catalog and feature-mask tooling are distinctive contributions. Cost/benefit: the dependency surface (many cloud SDKs) and learning cost are sizable for non-Azure users, hence a deduction.
Claim traceability is strong: every README feature links to a concrete example or package; the telemetry doc links a pinned .NET registry commit. Cross-source corroboration: README and feature-usage-telemetry.md agree on telemetry behavior, but static review cannot execute checks and SKILL-referenced remote gh-aw content is not in the packet, so 2. Fact/inference separation is good: implemented vs. unimplemented, preview status, and an explicit note that declaration matches do not establish behavioral parity — evidence is honest.
- The project is in public preview; APIs may change, so plan for churn in production use.
- Rollback documentation is thin: workflows support checkpointing, but framework-level recovery paths need independent verification.
- Feature-mask telemetry is on by default (Azure allowlisted hosts only); set AGENT_FRAMEWORK_FEATURE_MASK_DISABLED or USER_AGENT_DISABLED if no telemetry is desired.
- Data flows and risks for third-party systems (OpenAI, MCP, A2A, etc.) are the user's responsibility; audit data boundaries and approval mechanisms yourself.
- The go.mod dependency surface is large and includes a pseudo-version dependency (AG-UI); re-verify dependency provenance and vulnerability status for production.
FAQ
Can I use this in production?
What does it cost?
What telemetry does it collect?
What features are missing from the Go SDK?
Which model providers are supported?
Is Microsoft Agent Framework for Go open source?
Is Microsoft Agent Framework for Go safe to use?
Is Microsoft Agent Framework for Go still maintained?
Who makes Microsoft Agent Framework for Go?
What is Microsoft Agent Framework for Go's official website?
Thanks — we use this to improve the page.