agentfootprint
Find the context that made your agent answer wrong.
The explainable agent framework. Every run records its own causal trace — context, decisions, tool calls — as connected evidence. So when an answer is wrong, you backtrack from the output to the exact piece of context that caused it, confirmed by re-running without it.
Why is a query, not a guess.
It's built on footprintjs, runs
mock-first (no API key, $0) for local development, and integrates Anthropic,
OpenAI, and Bedrock when you're ready to ship.
The shape of it
- 2 primitives —
LLMCall(ask the model once) andAgent(the ReAct loop). - 4 compositions —
Sequence,Parallel,Conditional,Loop. - 1 Injection primitive — route facts, instructions, skills, and retrieval into the system / messages / tools slots of each call.
- 1 Memory factory — types × strategies, including Causal Memory (decision-evidence snapshots → cross-run replay and cheap-model triage).
- Typed observability events across many domains — every step is a recorded fact.
Start here
Getting started
Your first agent in a few lines — mock-first, no API key.
Why is a query, not a guess
The thesis: backtrack a wrong answer to its cause, with evidence.
API Reference
The full typed surface, auto-generated from source.
GitHub
Source, examples, and the issue tracker.
This site is the Fumadocs pilot for agentfootprint. The live Starlight docs remain at their current URL until the pilot is complete.