Answer in 50 Words
AutoGen entered maintenance in Oct 2025; Microsoft Agent Framework 1.0 (AutoGen + Semantic Kernel merge) went GA in Apr 2026 with YAML agent definitions, graph workflows with checkpointing, and native MCP + A2A support. I migrated a live triage repo on a ₹6K VPS in three evenings — 220 lines to 40, behavior preserved, bill flat.
InfoQ's Aug 2026 GA coverage Gujarat founders mostly missed: the successor is not a patch, it is a merged framework with staying power. New builds should never start on AutoGen now. Live repos face a real choice — migrate or wrap. I did both across two clients: migrated the active triage repo, wrapped the frozen reporting repo in MCP untouched. Migration notes below are the three evenings compressed.
War Story 1: The 220-Line Triage That Became 40 Lines YAML
The repo: vendor-risk triage, three roles, two tools, one HITL pause before sending recommendations back to the contract system. AutoGen wiring sprawled — two hundred twenty lines of agent setup, most of it plumbing. MAF port: forty lines of YAML definitions plus the same tool functions untouched, checkpointing preserved, MCP tools attached natively instead of through adapters. Token profile flat across the eval set. The evening-by-evening split: night one scaffolding + YAML definitions compiling, night two checkpoint + HITL pause verified with kill-and-resume tests, night three MCP/A2A attach plus full eval replay. Friday deploy, Monday silence — the good kind.
Migrate vs Wrap (Decision Table)
| Situation | Call | Why |
|---|---|---|
| Active repo, weekly changes | Migrate to MAF 1.0 | Maintenance mode means no fixes coming; YAML cuts future cost |
| Frozen profitable repo | Wrap in MCP, do not rewrite | Rewriting profit is vandalism; MCP front extends life years |
| .NET-heavy team | Migrate eagerly | MAF is the native Microsoft path, Semantic Kernel included |
| Python-only, LangGraph already | Wrap or re-platform | If LangGraph checkpoints already serve, migration is optional |
| Regulated HITL flows | Migrate with pause tests | Verify human-approval gates survive the port explicitly |
My .NET client migrated eagerly and gained. My Python reporting client stayed wrapped and saved a sprint. Both correct. Framework loyalty is not a virtue. Ledger outcomes are.
Code: Before and After (Runnable Shapes)
# before/autogen_triage.py — legacy shape (maintenance mode, do not extend)
# Agents defined imperatively; adapter needed for MCP tools
agents = {
"drafter": {"model": "gpt-5-6-sol", "tools": ["vendor_lookup"]},
"reviewer": {"model": "claude-mythos-5", "needs_approval": True},
}
# ~220 lines of setup, callbacks, and glue in the real file
# after/maf-triage.yaml — MAF 1.0 definitions (40 lines that matter)
agents:
- name: drafter
model: gpt-5-6-sol
tools: [mcp://catalog/vendor_lookup]
- name: reviewer
model: claude-mythos-5
requiresApproval: true
workflow:
graph:
- drafter -> reviewer
- reviewer.approved -> contract_system
- reviewer.rejected -> drafter
checkpointing: true
protocols: [mcp, a2a]
# after/verify_port.py — kill-and-resume + eval replay (run before deploy)
CASES = ["vendor_lookup basic", "hitl pause fires", "resume after kill", "a2a handoff shape"]
def verify(results: dict) -> bool:
must = ["hitl pause fires", "resume after kill"]
return all(results.get(c) == "pass" for c in must) and sum(1 for c in CASES if results.get(c) == "pass") >= 3
Don't do this: porting prompts verbatim without re-testing pause behavior. My first pass kept wording, lost the approval halt on one path — caught by the kill-and-resume test on night two, not by reading. Run the halt test. Trust the halt test.
When NOT to Migrate Yet
Frozen revenue repos, teams mid-festival with zero slack, and Python shops already checkpointed on LangGraph gain little this quarter. Wrap with MCP, schedule the port for the calm month, spend the sprint on evals instead. Migration is a cost with returns — only pay it where returns live. My wrapped reporting repo earns monthly untouched. Pride would have rewritten it. Arithmetic kept it.
War Story 2: The Halt Test That Earned Its Evening
Night two, 23:40, kill test: SIGKILL the triage mid-draft, resume from checkpoint, confirm the reviewer still demanded approval before the contract callback. First attempt resumed but skipped the pause — a ported flag defaulting wrong. Eleven minutes to find, one line to fix, full replay green by 00:20. That single test is why the Friday deploy stayed boring. Checkpoints without verified halts are decoration. Verify halts like production depends on it, because for procurement and finance flows it does.
Cost Ledger (Three Evenings)
| Line | Amount |
|---|---|
| Engineering (my own evenings) | 3 × 3h, unbilled R&D |
| Extra tokens (eval replays) | ~₹900 |
| Infra delta | ₹0 — same ₹6,200 VPS |
| Lines changed | 220 imperative → 40 YAML + tools untouched |
| Behavior delta | None on evals; halt verified live |
Payback: every future change edits YAML instead of plumbing. First change request (new approval rule) took twenty minutes instead of two hours. Migration paid for itself inside a month, and every YAML edit since has stayed boring in the best possible way.
Hiring Note: What a Migration Quote Should Contain
If you outsource this port, demand line items: YAML definition count, checkpoint plus halt verification runs named, MCP/A2A attach list, eval replay deltas before and after, and rollback commit hash. My fixed band for a triage-sized port is ₹35K–₹60K including the eval replay and one month of halt monitoring. Anyone quoting without naming the halt test is guessing. Anyone promising zero behavior delta without an eval set is hoping. Fixed price, named tests, ledger screenshot at handover — that is the whole procurement policy.
Frequently Asked Questions
Is AutoGen really dead in 2026?
AutoGen entered maintenance Oct 2025; Microsoft Agent Framework 1.0, merging AutoGen with Semantic Kernel, reached GA Apr 2026 per InfoQ. No new features land on AutoGen proper — start new builds on MAF or LangGraph, never fresh AutoGen.
How long does AutoGen to MAF migration take?
My live triage port took three evenings: YAML scaffolding, checkpoint plus HITL verification, MCP/A2A attach with eval replay. Frozen repos should wrap in MCP instead — zero evenings, years of extended life.
Does MAF support MCP and agent-to-agent protocols?
Yes — native MCP and A2A support ships in 1.0, which is the core reason to migrate active repos rather than maintain adapters. My port attached catalog tools as mcp:// endpoints with zero glue code.
How much does an AutoGen to MAF migration cost in India?
A triage-sized port runs ₹35K–₹60K fixed in my Junagadh band: YAML definitions, checkpoint plus halt verification, MCP/A2A attach, eval replay with deltas, rollback hash, and one month of halt monitoring. Frozen repos cost ₹0 to wrap in MCP instead — rewrite only what earns.
Should Python teams adopt a Microsoft framework?
Only where it pays: .NET adjacency, enterprise SSO needs, or active AutoGen repos. Python shops happily checkpointed on LangGraph can wrap and wait. I run both patterns behind one OPA gate and one ledger without drama.
Bottom Line
Migrate the living, wrap the profitable, verify halts like money depends on it. Three evenings, forty lines of YAML, same VPS, calmer future. Maintenance mode is a message — read it early, move on schedule.
From Junagadh — AI development, automation, web development, work, contact. Related: /journal/langgraph-deep-agents-vs-crewai-autogen-2026, /journal/state-of-ai-agents-252-tools-sep-2026.