With the ecosystem of agentic instruments and frameworks exploding in measurement, navigating the numerous choices for constructing AI programs is turning into more and more troublesome, leaving builders confused and paralyzed when choosing the proper instruments and fashions for his or her purposes.
In a new research, researchers from a number of establishments current a complete framework to untangle this advanced net. They categorize agentic frameworks primarily based on their space of focus and tradeoffs, offering a sensible information for builders to decide on the best instruments and techniques for his or her purposes.
For enterprise groups, this reframes agentic AI from a model-selection drawback into an architectural determination about the place to spend coaching price range, how a lot modularity to protect, and what tradeoffs they’re prepared to make between price, flexibility, and danger.
Agent vs. instrument adaptation
The researchers divide the panorama into two major dimensions: agent adaptation and instrument adaptation.
Agent adaptation entails modifying the inspiration mannequin that underlies the agentic system. That is finished by updating the agent’s inside parameters or insurance policies by way of strategies like fine-tuning or reinforcement studying to higher align with particular duties.
Software adaptation, however, shifts the main focus to the surroundings surrounding the agent. As an alternative of retraining the big, costly basis mannequin, builders optimize the exterior instruments corresponding to search retrievers, reminiscence modules, or sub-agents. On this technique, the primary agent stays “frozen” (unchanged). This method permits the system to evolve with out the large computational price of retraining the core mannequin.
The research additional breaks these down into 4 distinct methods:
A1: Software execution signaled: On this technique, the agent learns by doing. It’s optimized utilizing verifiable suggestions instantly from a instrument’s execution, corresponding to a code compiler interacting with a script or a database returning search outcomes. This teaches the agent the “mechanics” of utilizing a instrument accurately.
A major instance is DeepSeek-R1, the place the mannequin was skilled by way of reinforcement studying with verifiable rewards to generate code that efficiently executes in a sandbox. The suggestions sign is binary and goal (did the code run, or did it crash?). This technique builds robust low-level competence in steady, verifiable domains like coding or SQL.
A2: Agent output Signaled: Right here, the agent is optimized primarily based on the standard of its remaining reply, whatever the intermediate steps and variety of instrument calls it makes. This teaches the agent methods to orchestrate numerous instruments to achieve an accurate conclusion.
An instance is Search-R1, an agent that performs multi-step retrieval to reply questions. The mannequin receives a reward provided that the ultimate reply is right, implicitly forcing it to study higher search and reasoning methods to maximise that reward. A2 is right for system-level orchestration, enabling brokers to deal with advanced workflows.
T1: Agent-agnostic: On this class, instruments are skilled independently on broad information after which “plugged in” to a frozen agent. Consider traditional dense retrievers utilized in RAG programs. A normal retriever mannequin is skilled on generic search information. A strong frozen LLM can use this retriever to seek out data, although the retriever wasn’t designed particularly for that LLM.
T2: Agent-supervised: This technique entails coaching instruments particularly to serve a frozen agent. The supervision sign comes from the agent’s personal output, making a symbiotic relationship the place the instrument learns to offer precisely what the agent wants.
For instance, the s3 framework trains a small “searcher” mannequin to retrieve paperwork. This small mannequin is rewarded primarily based on whether or not a frozen “reasoner” (a big LLM) can reply the query accurately utilizing these paperwork. The instrument successfully adapts to fill the particular information gaps of the primary agent.
Complicated AI programs would possibly use a mixture of those adaptation paradigms. For instance, a deep analysis system would possibly make use of T1-style retrieval instruments (pre-trained dense retrievers), T2-style adaptive search brokers (skilled through frozen LLM suggestions), and A1-style reasoning brokers (fine-tuned with execution suggestions) in a broader orchestrated system.
The hidden prices and tradeoffs
For enterprise decision-makers, selecting between these methods usually comes down to 3 elements: price, generalization, and modularity.
Price vs. flexibility: Agent adaptation (A1/A2) provides most flexibility since you are rewiring the agent’s mind. Nevertheless, the prices are steep. For example, Search-R1 (an A2 system) required coaching on 170,000 examples to internalize search capabilities. This requires large compute and specialised datasets. Then again, the fashions may be rather more environment friendly at inference time as a result of they’re much smaller than generalist fashions.
In distinction, Software adaptation (T1/T2) is way extra environment friendly. The s3 system (T2) skilled a light-weight searcher utilizing solely 2,400 examples (roughly 70 occasions much less information than Search-R1) whereas reaching comparable efficiency. By optimizing the ecosystem somewhat than the agent, enterprises can obtain excessive efficiency at a decrease price. Nevertheless, this comes with an overhead price inference time since s3 requires coordination with a bigger mannequin.
Generalization: A1 and A2 strategies danger “overfitting,” the place an agent turns into so specialised in a single activity that it loses common capabilities. The research discovered that whereas Search-R1 excelled at its coaching duties, it struggled with specialised medical QA, reaching solely 71.8% accuracy. This isn’t an issue when your agent is designed to carry out a really particular set of duties.
Conversely, the s3 system (T2), which used a general-purpose frozen agent assisted by a skilled instrument, generalized higher, reaching 76.6% accuracy on the identical medical duties. The frozen agent retained its broad world information, whereas the instrument dealt with the particular retrieval mechanics. Nevertheless, T1/T2 programs depend on the information of the frozen agent, and if the underlying mannequin can’t deal with the particular activity, they are going to be ineffective.
Modularity: T1/T2 methods allow “hot-swapping.” You’ll be able to improve a reminiscence module or a searcher with out touching the core reasoning engine. For instance, Memento optimizes a reminiscence module to retrieve previous instances; if necessities change, you replace the module, not the planner.
A1 and A2 programs are monolithic. Instructing an agent a brand new ability (like coding) through fine-tuning could cause “catastrophic forgetting,” the place it degrades on beforehand discovered abilities (like math) as a result of its inside weights are overwritten.
A strategic framework for enterprise adoption
Based mostly on the research, builders ought to view these methods as a progressive ladder, shifting from low-risk, modular options to high-resource customization.
Begin with T1 (agent-agnostic instruments): Equip a frozen, highly effective mannequin (like Gemini or Claude) with off-the-shelf instruments corresponding to a dense retriever or an MCP connector. This requires zero coaching and is ideal for prototyping and common purposes. It’s the low-hanging fruit that may take you very far for many duties.
Transfer to T2 (agent-supervised instruments): If the agent struggles to make use of generic instruments, do not retrain the primary mannequin. As an alternative, prepare a small, specialised sub-agent (like a searcher or reminiscence supervisor) to filter and format information precisely how the primary agent likes it. That is extremely data-efficient and appropriate for proprietary enterprise information and purposes which are high-volume and cost-sensitive.
Use A1 (instrument execution signaled) for specialization: If the agent basically fails at technical duties (e.g., writing non-functional code or incorrect API calls) you will need to rewire its understanding of the instrument’s “mechanics.” A1 is finest for creating specialists in verifiable domains like SQL or Python or your proprietary instruments. For instance, you possibly can optimize a small mannequin to your particular toolset after which use it as a T1 plugin for a generalist mannequin.
Reserve A2 (agent output signaled) because the “nuclear choice”: Solely prepare a monolithic agent end-to-end when you want it to internalize advanced technique and self-correction. That is resource-intensive and barely essential for traditional enterprise purposes. In actuality, you not often have to become involved in coaching your individual mannequin.
Because the AI panorama matures, the main focus is shifting from constructing one large, good mannequin to developing a wise ecosystem of specialised instruments round a steady core. For many enterprises, the simplest path to agentic AI is not constructing a much bigger mind however giving the mind higher instruments.
