MiniMax Agent: What We Learned While Building in 2025

                                          Practical discoveries from real-world agent development

Over the past year, we’ve shifted from chatbots built on stacked engineering pipelines to Agents that orchestrate tools and sub-agents to solve real, complex, real-world problems. This transition forced us to fundamentally rethink how we design, build, and ship AI products.
Here are five key takeaways from our team’s journey in 2025👇

SOTA models are evolving faster than any fixed system can keep up with. Instead of force-fitting new models into existing pipelines, we’ve learned to tinker first —to explore each model’s native strengths before committing to architecture.
For example, when rethinking our web-editing feature, strong VLMs allowed us to scrap entire chains of brittle logic. Spatial reasoning meant we could simply “paint” edits directly onto a page. Similarly, experimenting with image-to-slides models replaced our legacy HTML-conversion flow and unlocked a far more flexible presentation experience in custom modes.
The lesson: let models show you what they’re good at before telling them what to do.

Not every workflow needs an Agent, and that’s okay.
For many existing systems, simply replacing a manual step with an LLM node delivers massive value. Deterministic workflows still dominate low-entropy, well-defined tasks. The Agent Threshold is crossed when problems become open-ended, ambiguous, or too complex for predefined paths. That’s when we hand control to the model.
Our approach is to let Agents “travel light”: start with a minimal system prompt and core knowledge, then iteratively tighten behavioral boundaries based on how the Agent navigates real-world uncertainty and decision-making.

The line between Product, Design, and Engineering is rapidly dissolving—everyone is a builder now.
We’ve moved away from static PRDs toward interactive demos, defining products by their vibe and the first 10 user queries rather than long feature lists. With tools like Cursor and strong VLMs (e.g. Gemini 3 Pro), we can prototype complete user journeys in hours.
This lets us align on the actual outcome much faster—validating product feel before writing a single line of production code.

Most real-world tasks (SEO strategy, growth planning, ops analysis) don’t have a single ground truth like academic benchmarks.
We evaluate Agents the way we’d evaluate junior employees: focusing on outcome quality, reasoning trajectory, and consistency rather than rigid scores. Our “Agent-as-Judge” setups involve evaluator Agents that actually run code, verify data, or track results over time.
The goal isn’t perfection but low variance. A reliable digital colleague should deliver stable, predictable performance with a reasonable ROI.

Tools alone aren’t enough. Professional-grade delivery requires more context, memory, and triggers.
We’re moving away from one-off executions toward Agents that live inside their environments. A Marketing Agent, for example, shouldn’t just run a search—it should monitor trends, trigger workflows on traffic spikes, and reflect on performance data over time.
This persistent, context-rich operation is the only way to approximate how real professionals actually work.

We’re excited about Generative UI—interfaces that are fluid and rendered in real time by an Agent’s reasoning, rather than constrained by predefined controls.
At the same time, we’re pushing Agents deeper into professional digital environments. As we integrate vertical knowledge (Legal, Audit, Finance, and beyond), Agents will move from executing tasks to internalizing professional logic.
If you have expert workflows worth encoding, contribute them here 👉 https://minimax-contributor-program.space.minimax.io/

See you in the next iteration!