Agents
GenAIScript defines an agent as a tool that runs an inline prompt to accomplish a task. The agent’s LLM is typically augmented with additional tools and a memory.
GenAIScript does not implement any agentic workflow or decision. It relies entirely on tools support built into the LLMs.
Agent = LLM + Tools
Let’s take a look at the agent_git
example that query a git repository. This agent is registered as a tool
and can be used in the LLM prompt.
When the LLM needs information about something like “summarize changes in the current branch”, it will call the agent_git
tool with the query get changes in the current branch
.
The agent_git
tool itself has access to various git dedicated tools like git branch
, git diff
that it can use to solve.
It will have to resolve the current and default branch, compute a diff and return it to the main LLM.
Agent vs Tools
Note that in this simple example, you could also decide to flatten this tree and give access to the git tools to the main LLM and skip the agent.
However, the agent abstraction becomes useful when you start to have too many functions or to keep the chat conversation length small as each agent LLM call gets “compressed” to the agent response.
Multiple Agents
Let’s take a look at a more complex example where multiple agents are involved in the conversation. In this case, we would like to investigate why a GitHub action failed.
It involves the agent_git
and the agent_github
agents. The agent_github
can query workflows, runs, jobs, logs and the agent_git
can query the git repository.
Memory
All agents are equipped with a memory that allows them to share information horizontally across all conversations.
The memory is a log that stores all agent / query / answer
interactions. When generating the prompt for an agent,
the memory is first prompted (using a small LLM) to extract relevant information
and that information is passed to the agent query.
All agents contribute to the conversation memory unless it is explicitly disabled using disableMemory
.
defAgent
The defAgent
function is used to define an agent that can be called by the LLM. It takes a JSON schema to define the input and expects a string output. The LLM autonomously decides to call this agent.
- the agent id will become the tool id
agent_<id>
- the description of the agent will automatically be augmented with information about the available tools
Builtin Agents
Example agent_github
Let’s illustrate this by building a GitHub agent. The agent is a tool that receives a query and executes an LLM prompt with GitHub-related tools.
The definition of the agent looks like this:
and internally it is expanded to the following:
Inside callback, we use runPrompt
to run an LLM query.
- the prompt takes the query argument and tells the LLM how to handle it.
- note the use of
ctx.
for nested prompts
Selecting the Tools and System Prompts
We use the system
parameter to configure the tools exposed to the LLM. In this case, we expose the GitHub tools (system.github_files
, system.github_issues
, …)
This full source of this agent is defined in the system.agent_github system prompt.