Chinese AI lab MiniMax released M2.7, a model it describes as “deeply participating in its own evolution.” The release is notable less for its benchmark numbers and more for what it reveals about how the company is using agent harnesses as a training methodology, not just a deployment pattern.
The Self-Evolution Loop
MiniMax tasked an internal version of M2.7 with building a research agent harness that interacts with different project groups across the company. The harness handles data pipelines, training environments, infrastructure, cross-team collaboration, and persistent memory. Researchers then drive the harness to produce better models.
In practice, according to MiniMax’s announcement, the daily workflow looks like this: a researcher discusses an experimental idea with the agent, which handles literature review, experiment specification, data pipelining, and launch. During experiments, the agent monitors progress and automatically triggers log reading, debugging, metric analysis, code fixes, merge requests, and smoke tests. Human researchers only intervene for critical decisions.
MiniMax claims the model handles 30-50% of the RL team’s workflow autonomously.
Recursive Harness Improvement
The more interesting claim is recursive self-improvement of the agent infrastructure itself. MiniMax says M2.7 ran entirely autonomously through over 100 rounds of an iterative loop: analyze failure trajectories, plan changes, modify scaffold code, run evaluations, compare results, decide to keep or revert changes.
The internal harness “autonomously collects feedback, builds evaluation sets for internal tasks, and based on this continuously iterates its own architecture, skills/MCP implementation, and memory mechanisms,” according to the release post.
Benchmarks
On standard evaluations, M2.7 scored 56.22% on SWE-Pro (approaching Opus-level performance), 55.6% on VIBE-Pro for end-to-end project delivery, and 57.0% on Terminal Bench 2 for complex engineering systems. The model maintains a 97% skill adherence rate while working with over 40 complex skills, each exceeding 2,000 tokens.
Its GDPval-AA ELO score of 1495 is, per MiniMax, the highest among open-source models for professional office software tasks including complex Excel, PowerPoint, and Word editing.
Entertainment Agents via OpenRoom
Beyond productivity, MiniMax positioned M2.7 as a platform for character-consistent interactive agents. The model shows “excellent character consistency and emotional intelligence,” and the company launched OpenRoom as an interactive agent system for entertainment applications.
This is a deliberate differentiation play. While OpenAI, Anthropic, and Google compete on coding agents and enterprise productivity, MiniMax is placing a bet that the larger market for persistent AI agents may be entertainment and companionship rather than software engineering.
The Agentic Training Frontier
The release signals a broader shift. When a model actively builds the infrastructure used to train its successor, the boundary between tool and toolmaker blurs. MiniMax is not the first to explore self-play and recursive improvement, but publicizing a concrete workflow where the model handles half the RL team’s work and iterates its own scaffold over 100+ autonomous rounds sets a specific benchmark for what “agentic training” looks like in production.