Can agents discover tools dynamically instead of pre-selecting them?
Explore whether agents can find needed tools during execution rather than choosing from a fixed set upfront. This matters for long-horizon tasks where relevant tools cannot be known in advance.
Most current agent frameworks — ReAct, Plan-and-Solve, and their variants — work with predefined tool sets. Before execution starts, the agent is given a fixed set of tools it can use. The agent then operates within that set: reason about the task, pick a tool, observe the result, reason again, pick another tool. This works well for tasks where the relevant tools are known in advance.
For long-horizon tasks where the relevant tools cannot be enumerated up front, the predefined approach hits limits. The agent does not know early in the task what it will need later. A tool that becomes essential mid-task may not have been included in the initial set. Adding "all possibly relevant tools" up front bloats the context and confuses tool selection.
DeepAgent takes a different architectural stance: dynamic tool discovery during execution. Tools are not pre-retrieved; they are discovered on an as-needed basis as the agent progresses through the task. The agent maintains a global perspective on the entire task and decides when to search for new tools based on what the current state of execution actually requires.
The shift has consequences for how the agent reasons. ReAct-style workflows force per-step deliberation on specific isolated operations — "what should I do this turn?" — which can lose the global picture. Dynamic discovery preserves the global picture: the agent is not constrained to deliberate on the next operation alone but can reconsider the whole task arc, including which tools to bring into the picture next.
This connects to memory folding (same paper). Memory folding lets the agent pause to reconsider strategies; dynamic tool discovery is what makes the reconsidered strategy actionable — the agent can fetch new tools after the reconsideration. The two mechanisms together produce an agent that is not locked into either a fixed plan or a fixed toolset.
For long-horizon agent deployments — research assistants, multi-step software engineering, complex workflow automation — dynamic tool discovery beats pre-retrieval for tasks where the relevant tool space is too large to enumerate.
Related concepts in this collection
-
Can agents compress their own memory without losing critical details?
Explores whether agents can autonomously consolidate interaction history into structured memory schemas that reduce token overhead while preserving information needed for long-horizon reasoning and strategic reflection.
same paper, the memory mechanism that pairs with this workflow
-
Can simulated APIs and token-level credit assignment train better tool-using agents?
Training agents to use real APIs is expensive and unstable, and sparse rewards make it hard to credit the right tool calls. Can combining LLM simulators with fine-grained advantage attribution solve both problems?
same paper, the training method that supports the workflow
-
Can semantic capability vectors replace manual agent routing?
Explores whether embedding agent capabilities in high-dimensional space and matching them semantically can eliminate brittle, manually-maintained topic-based routing in multi-agent systems.
adjacent: dynamic capability discovery in multi-agent systems
Click a node to walk · click center to open · click Open in graph to see this note in the full knowledge graph
Original note title
dynamic tool discovery during execution beats pre-retrieved tool sets for long-horizon agents — global task perspective unconstrained by predefined ReAct workflows