跪拜 Guibai
← All articles
Artificial Intelligence · Agent · LLM

Stop Comparing Agent Frameworks — First Decide Whether You Need a Product, a Chassis, or a Workflow

By 武子康 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Agent tooling has split into three distinct layers — product, chassis, and workflow runtime — and teams that compare them on a single feature axis end up buying the wrong abstraction. The cost shows up as unowned state, duplicated governance, or a runtime too heavy for the actual job.

Summary

Pitting DeepSeek Harness against Codex, Claude Code, or LangGraph in a feature checklist conflates three distinct procurement layers. Codex and Claude Code are coding agent products that ship with a working terminal, repository workflow, and approval experience. DeepSeek Harness is a developer-preview agent runtime where the loop, session, tool registry, and surface all live inside a reversible plugin graph. LangGraph is a low-level orchestration runtime that forces teams to model business steps, state, and recovery points as explicit graphs with checkpoints and interrupts.

A team that needs delivery speed should buy a product. A team building multiple agent surfaces on shared primitives should evaluate Harness. A team whose core problem is pausable, auditable business logic should reach for LangGraph. The wrong choice loads the team with complexity it never needed, while the right one leaves a clear acceptance checklist for whatever the system does not handle.

No success-rate, cost, or latency ranking holds without a controlled experiment that fixes the model, task set, permissions, sandbox, and scoring rubric. The adoption matrix here is a starting filter, not a substitute for a same-conditions PoC.

Takeaways
Codex and Claude Code are coding agent products that already own the terminal, repository workflow, and approval loop; a team that only needs a working assistant should start here.
DeepSeek Harness exposes the agent loop, session, tool registry, and surface inside a single Cordis plugin graph, making it a chassis for teams building multiple agent products on shared primitives.
LangGraph forces explicit state, checkpoints, and interrupts, which suits approval flows and cross-system business processes where every step must be pausable and recoverable.
Calling something a “plugin” means different things across these systems: Claude Code plugins extend a product lifecycle, while Harness plugins can replace core runtime layers.
Combining two runtimes (e.g., LangGraph for outer state and Harness for inner agent nodes) requires answering who owns the task ID, cancellation propagation, checkpoint idempotency, and trace correlation.
A feature-table comparison of tools, subagents, and persistence is useless without a controlled experiment that fixes the model, task set, permissions, sandbox, and scoring rubric.
The adoption matrix is a validation filter, not a final decision; every choice still demands a same-conditions PoC before production commitment.
Conclusions

Most “which agent framework is better” debates collapse because the participants are buying different things: a product, a runtime chassis, or a state-machine library. The feature lists overlap, but the ownership boundaries do not.

DeepSeek Harness’s reversible plugin lifecycle is architecturally unusual — effects unwind on unload and dependency changes re-trigger consumers — which shifts it from a coding tool into a platform-building substrate.

Claude Code’s extension model ties plugins, skills, agents, and hooks to a product lifecycle, not a runtime core; this makes it easier to distribute team assets but harder to replace the underlying agent loop.

LangGraph’s interrupt-and-resume semantics force a hard question most agent loops dodge: if a node re-executes from the top on resume, every side effect before the interrupt must be idempotent, and that burden falls entirely on the developer.

Combining two runtimes without first defining ID ownership, cancellation propagation, and checkpoint idempotency turns a recoverable system into a dual state machine with no single source of truth for task completion.

Concepts & terms
Cordis Plugin Graph
DeepSeek Harness’s composition model where plugins declare effects and dependencies; effects can be reversed on unload, and dependency changes trigger consumer re-execution, making the entire agent runtime recomposable.
Checkpoint (LangGraph)
A persisted snapshot of a thread’s graph state, saved by a Checkpointer. Combined with thread_id and interrupt(), it enables pausing execution, waiting for external input, and resuming — but the node re-executes from the start, requiring idempotent side effects.
Append-only Session Event
DeepSeek Harness’s logging model where raw events are only appended, never rewritten. A separate derived Surface is what the model sees; compaction works by replacing the visible interval, not by mutating history.
Subagent Context Window
In Claude Code, a subagent runs in its own isolated context window with restricted tools and permissions, allowing separation of concerns — e.g., one subagent searches logs while another implements code — without polluting the main session.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗