Researchers propose a lighter, modular way to build AI agents
A new study introduces AgentForge, a small open-source toolkit for building AI programs that can plan and carry out multi-step tasks. The authors say today’s popular tools are often heavy, hard to modify and tie users to one provider. Their approach aims to make these systems easier to build, switch and supervise.
Why this is in the news
Interest in task‑doing AI has grown quickly as large language models (text‑generating AI systems) have improved. The paper, released on the open research platform arXiv by researchers Akbar Anbar Jafari, Cagri Ozcinar and Gholamreza Anbarjafari, argues that the next step is better engineering, not bigger models.
Why many AI tools feel rigid
The authors point to a structural issue: frameworks that bundle many features often become complex, slow to adapt and dependent on a single cloud service. That makes it costly to prototype, hard to audit how decisions are made and risky to deploy at scale.
A concrete example
Consider an assistant asked to compare prices from several shops, summarise the findings and send a short email. AgentForge breaks this into small, reusable "skills" with clear inputs and outputs (for example: "fetch page", "extract price", "compare", "draft email"). The skills are connected like a simple flowchart, so the process can run steps in order or in parallel, and each step can be tested on its own.
Key risk: speed and scale
The authors see the main risk in building bigger systems on top of tangled, provider‑locked code. When tasks run fast and at scale, small errors spread quickly and are hard to trace. If every change requires touching many parts, teams ship slower and trust the results less.
What they propose
AgentForge offers three practical measures. First, a modular skill system with clear contracts keeps tasks small and testable. Second, a unified interface lets teams switch between cloud services and local models without changing the task logic. Third, plain‑text settings files separate what an agent should do from how it is implemented, improving oversight.
In tests across four scenarios, the authors report similar task success to well‑known toolkits while cutting development time by 62% compared to LangChain and by 78% compared to writing directly against model APIs. The coordination adds less than 100 milliseconds of delay, which the team says is suitable for real‑time use. The framework ships with six built‑in skills and documentation for adding more.
Bottom line
If adopted, this design could make AI assistants more predictable, easier to audit and less tied to one provider. It is a technical paper, but the goal is straightforward: simpler parts, clearer connections and controls that make scaling safer.
In a nutshell
AgentForge is a small, open toolkit that breaks AI tasks into clear, swappable parts, aiming to cut build time, avoid lock‑in and keep systems easier to control.
What to take away
- Small, testable building blocks make complex AI tasks easier to trust and maintain.
- Being able to switch between cloud and local models reduces dependence on any single provider.
- Clear separation between task logic and settings improves oversight and speeds up changes.
Paper: https://arxiv.org/abs/2601.13383v1
Register: https://www.AiFeta.com
AI research opensource software agents languagemodels
