Deep Investment Research18 min read
DeepSeek Harness Reached 68,600 Stars in a Day. The Most Valuable AI Asset May No Longer Be the Model
DeepSeek open-sourced an MIT-licensed agent runtime where the model, tools, log and loop are replaceable. It may commoditize agent software while creating a harder-to-copy moat in workflow trajectories.

The Operating Layer for Agents
The MIT-licensed project makes the model adapter, tools, session log and agent loop replaceable. It supports competing models and can delegate to Claude Code and Codex.
The harness increasingly owns the steering wheel, permissions and driving record while the model becomes a swappable engine.
Benchmarks Belong to Systems
DeepSeek reports V4-Pro at 87.9 on Terminal Bench, 62.7 on DeepSWE and 74.1 on Toolathlon. Prompting, context, retries and verification all contribute, so agent scores cannot be assigned to the model alone.
The Data Asset Behind Free Software
An append-only event stream records which files the agent saw, tools it called, failures it made and fixes that passed. These outcome-labeled trajectories are more valuable for post-training than final chat text.
Local deployment does not automatically return data to DeepSeek; monetization still depends on hosted services, telemetry and API share.

The Investment Test
Claude Code had more than $2.5B in run-rate revenue, proving the harness can monetize. But stars are not cash flow, and model neutrality can route usage away from DeepSeek.
Use a reverse DCF to test paid conversion and API revenue assumptions.
Recent posts








































































%20-1.jpg?width=1600)
Frequently asked questions
Does it only support DeepSeek?
No. It supports multiple model providers and Claude Code or Codex subagents.
What are the main risks?
Traffic leakage, plugin security, maintenance burden and weak monetization.
Value distribution, task cost, data flywheel and monetization separately.
Download Valuefocus