Agent framework / architecture review2026-08-30
From gateway to worker
Agent Framework
Comparison / 2026
Where AutoGen, LangChain/LangGraph, OpenClaw, Pi, DSH and Microsoft Agent Framework fit in a production system.
Role clarityStateful workflowsWorker contractsMigration gates
AI
×
SYS
02 / 08Decision frame
The first question is ownership
These projects are not the same kind of thing.
Compare the layer they own.
Gateway / control plane
Channels, sessions, routing, tenant policy, delivery and approvals.
HermesOpenClaw
→Orchestration
State, graph edges, messages, retries, handoffs and termination.
LangGraphAutoGenMAF
→Execution / coding worker
Workspace, tools, extensions, RPC, coding loop and trajectory.
PiDSH
Observability and evaluation
OpenTelemetry, LangSmith or another backend sees the run. It does not automatically own policy, tenant isolation or delivery.
A framework is relevant when it removes a real operating burden.
Popularity is not an architecture.
03 / 082026 status check
AutoGen is a migration concern; LangGraph is a current orchestration option
Still relevant?
Yes, but for different reasons.
AutoGen / official statusRelevant as an existing-system bridge.
Microsoft's repository now labels AutoGen as maintenance mode: no new features, community managed, and new users are directed to Microsoft Agent Framework.
- Keep it for an existing AutoGen app, benchmark or migration window.
- Do not make it the default foundation for a new 2026 production system.
- Its message-oriented team model remains useful for explaining collaboration and extraction of a worker contract.
Decision: migrate or contain. Do not expand the dependency.
Source: Microsoft / AutoGen repository
LangChain / LangGraph / official positioningRelevant when state is the product.
LangGraph is positioned as a low-level orchestration framework for long-running, stateful agents. LangChain remains the integration and model/tool layer around it.
- Use it for explicit state, branching, durable execution, human review and restartable workflows.
- Use LangChain for provider, retriever and tool integration when those abstractions save time.
- Keep gateway, tenant and delivery ownership outside the graph unless you intentionally choose the hosted platform.
Decision: a credible 2026 orchestration layer, not a complete product gateway.
Source: LangGraph overview
Microsoft Agent Framework is the current Microsoft successor path for new work: official repository. Status checked 2026-08-30.
04 / 08Role-based scorecard
Score the fit, not the feature list
Six options. Four ownership questions.
The winning stack is usually composable.
| Option | Control plane | Orchestration | Coding worker | 2026 fit | Use it when |
|---|
| Hermes → DSH | 5 / 5 | 4 / 5 | 4 / 5 | 5 / 5 | Current product assets and fastest production path matter. |
| OpenClaw + Pi | 5 / 5 | 4 / 5 | 4 / 5 | 4 / 5 | Greenfield product needs a thin gateway-to-runtime split. |
| OpenClaw + DSH | 5 / 5 | 4 / 5 | 5 / 5 | 3 / 5 | Benchmarks prove DSH beats Pi for long coding tasks or workspaces. |
| LangChain + LangGraph | 2 / 5 | 5 / 5 | 3 / 5 | 4 / 5 | Stateful workflow control and provider/tool portability are core. |
| Microsoft Agent Framework | 3 / 5 | 5 / 5 | 3 / 5 | 4 / 5 | New Microsoft-aligned Python/.NET multi-agent work needs a supported path. |
| AutoGen | 1 / 5 | 3 / 5 | 2 / 5 | 1 / 5 | An existing system already depends on it and migration has not started. |
Scale: 1 = weak fit for this role, 5 = strong fit.Architecture reference: DSH OpenClaw Architecture Comparison, 2026-08-17.
05 / 08Architecture choice
Keep the current production line intact while testing the next one
One production path.
One shadow path.
01 / CURRENT ASSETSHermes → DSH
Hermes owns channels, sessions, policy and delivery. DSH owns coding, workspace and trajectory.
Recommended production
02 / GREENFIELDOpenClaw + Pi
OpenClaw provides the gateway; Pi supplies the thin runtime. The split is light, but permissions, tenant policy, audit and quotas still need to be built.
Recommended shadow
03 / CONDITIONALOpenClaw + DSH
Higher coding ceiling, heavier adapter and lifecycle ownership. Choose only after a worker benchmark proves the gain.
Benchmark first
04 / PLATFORM BUILDDSH alone
Most consistent runtime story, but it also becomes UI, gateway, auth, tenant and delivery platform.
Not the first move
LangGraph, MAF and AutoGen sit inside this map as orchestration choices. None of them removes the need to name the gateway owner, worker owner or delivery owner.
06 / 08Migration boundary
Do not bind gateway migration to runtime migration
Define one contract.
Let the worker change behind it.
Contract fields
- Task passport
tenant / user / project / workspace / task - Worker events
start / progress / tool / approval / done - Policy owner
gateway or sidecar; worker never bypasses it - Session owner
child runtime manages its own local state - Delivery owner
gateway sends the final message and files - Evaluation fields
latency / cost / success / trajectory / retries
task = {
"tenant_id": tenant,
"workspace": workspace,
"task_id": task_id,
"worker": "dsh | pi | langgraph | maf"
}
event = {
"type": "tool | approval | done",
"evidence": evidence,
"next": next_action
}
Why it matters: the same contract can connect Hermes-native, DSH, Pi/OpenClaw+Pi, OpenClaw+DSH, LangGraph or MAF without moving the gateway, tenant model and delivery system together.
07 / 08Recommended route
A decision path with explicit gates
Production stays stable.
The shadow line earns its way in.
PRODUCTION NOWHermes Gateway → DSH worker
Reuse current channel, session, policy and delivery assets. Measure coding success, trajectory quality and operator time.
→
SHADOW NEXTOpenClaw Gateway → Pi worker
Run the same tasks and metrics. Test operational stability, session isolation and the thinness of the runtime boundary.
Gate 01 / QualityIs Pi close to DSH on the real coding workload?
Gate 02 / OperationsAre startup, provider env, events and plugin dependencies stable?
Gate 03 / EconomicsDoes migration benefit exceed adapter, support and retraining cost?
Use LangGraph when durable state and human checkpoints are the bottleneck. Use MAF for new Microsoft-aligned work. Keep AutoGen behind a migration boundary.
08 / 08Three sentences
Final answer
AutoGen is still useful.
It is no longer the default.
LangChain remains valuable as an ecosystem; LangGraph is the relevant orchestration layer for stateful agent work in 2026.
For this architecture, keep Hermes → DSH in production, validate OpenClaw + Pi in shadow, and switch only when the worker contract and benchmarks say the move is worth it.
01 / ROLEGateway, orchestration, worker and observability are separate decisions.
02 / STATUSAutoGen = maintenance and migration. LangGraph = active orchestration option.
03 / ROUTEHermes → DSH production. OpenClaw + Pi shadow.
04 / GATEBenchmark quality, operations and total migration cost before switching.
Source deck: DSH OpenClaw Architecture ComparisonOfficial project docs checked 2026-08-30