Harness Over Model: How AI Agents Evolved From Chat to Work in Five Stages
The same model tripled its benchmark score with zero changes to its weights, which puts agent performance squarely in the runtime layer. Open harnesses from OpenAI and DeepSeek mean any team can fork, audit, and embed that layer, so the deciding factors become runtime engineering, skill packaging, and acceptance-criteria discipline rather than which API you subscribe to.
The defining 2026 event in the agent industry came from OpenAI's August 20 open-sourcing of the Codex harness: harness-level optimization alone lifted GPT-5.6 Sol's ARC-AGI-3 from 13.3% to 38.3% and cut token consumption to one-sixth, with the model weights unchanged. DeepSeek had shipped its MIT-licensed harness a week earlier, built on the Cordis kernel where every component from the agent loop to the Web UI is a plugin. Both vendors now treat harness design, not raw model capability, as the primary lever on agent performance.
The five-stage evolution runs from prompt templates and ReAct through RAG engineering, coding assistants with compiler feedback, and conversational agents, to the current stage where agents are deployed, packaged, and traded as productivity units. The 2026-07-28 MCP spec landed a stateless core, async Tasks, and Skills over MCP; Agent Skills and Agent Plugins 1.0 standardized how capabilities get described and shipped. The comparison across five tools and eight model families puts Codex and DeepSeek Harness at the open-runtime poles, WorkBuddy at the multi-expert office pole, and Doubao at the all-modal real-time pole, while Kimi K3's 300-subagent swarm and MiniMax M3's 1M-context open weights mark the capability extremes.
Harness design has replaced raw model capability as the industry's primary performance lever; the same model tripling its benchmark score with unchanged weights makes the model-centric narrative of the last two years untenable.
The stateless MCP core won on a mundane operational ground: session-state glue capped integration scale at dozens of connections, and removing it lets enterprises reach thousands. The protocol's success tracks a cost curve, not a technical fashion.
DeepSeek's disclosures, a 13,467-token PONG reply and an unrestricted read/network sandbox, show open harnesses trade polish for transparency: the security burden moves to the adopter.
The human keeps moving one slot back in the loop across all five stages, but responsibility does not move with it; acceptance-criteria writing becomes the scarce skill. That reshapes hiring and job definitions as much as it reshapes tooling.
MiniMax's missing SOC 2, ISO 27001, and HIPAA credentials, alongside DeepSeek's procurement scrutiny, put a compliance ceiling on open-weight Chinese models in regulated Western markets that no benchmark score can remove.
Voice is the last modality standing between assembled and native all-modal systems: MiniMax's API-layer stitching of speech against Doubao's end-to-end SeedRealtime defines the near-term dividing line.