- Lightweight: Multi-agent framework based on a simple step model that provides natural crash-resume capability, and agent / user / env message coordination. Agents work through a small set of generic tools: shell, file edit, sub-agent spawn, and inter-agent coordination.
- Robust: DB-first persistence; crashing at any point during the agentic loop can be robustly resumed.
- The agentic loop is guaranteed to be consistent, failures in the tool / environment are not handled by the harness, but reported to the agent.
- Sub-agent session trees are automatically resumed upon server recovery.
- Autonomous: Each run can be fully driven by the autonomous agent, or cooperatively driven with user interactions.
Amplio is a single binary with an embedded frontend. To build it yourself you need Go (see go.mod for the version) and Node.js 22+:
make buildAmplio needs two things: a data directory (a local SQLite database and everything else it owns) and an LLM provider.
Write ~/.amplio/config.toml:
# For bookkeeping: run reports, summaries, compaction.
system_llm_hq = "vertex-gemini:gemini-3.1-pro-preview"
system_llm_fast = "vertex-gemini:gemini-3.6-flash"
# Enables skill + lesson search. Recommended.
embed_model = "vertex:text-embedding-005"
[run] # The model menu in the new-run form.
llms = [
"vertex-claude{cache_ttl=1h}:claude-opus-5?thinking.type=adaptive&thinking.display=summarized",
"vertex-gemini:gemini-3.6-flash",
]Please consult GCP prompt caching doc
before using the {cache_ttl=1h} control on the vertex-claude provider.
For Vertex AI, point ADC at your GCP project:
export VERTEXAI_PROJECT=<your-gcp-project>
export VERTEXAI_LOCATION=global
gcloud auth application-default loginThen start the server and open the URL it prints:
amplio serveFrom there, describe a task in the composer and start a run.
| Doc | Summary |
|---|---|
| Runs | autonomous and interactive runs, sub-agents, reports and grades |
| Configuration | the data directory, config.toml, precedence, recall |
| Models | LLM provider specs, thinking controls, bridges, embedders |
| Command line | serve, client, headless, and scripting them |
| What the agent knows | AGENTS.md, briefings, skills and lessons |
| Operations | multiple instances, access and tokens, backups, logs |
Design notes on how amplio works inside are in docs/internals.
Amplio does not sandbox the agent's shell: an agent runs commands as the user that started the server, with that user's filesystem access, network access and credentials. Run it somewhere you are comfortable with that.
Reads are also open to anyone who can reach the port — the token in the startup URL gates writes, not reads. See operations for what that means and how to bind it down.
This is not an officially supported Google product. This project is not eligible for the Google Open Source Software Vulnerability Rewards Program.
If you find Amplio helpful, please cite the following BibTeX:
@misc{Zhang2026Amplio,
author = {Chiyuan Zhang and Da Huang and Chen Liang and Andrew Li and {Amplio Contributors}},
title = {{Amplio: A Lightweight Agent Harness for Robust and Long-Horizon Runs}},
year = {2026},
howpublished = {GitHub repository},
url = {https://github.com/google-deepmind/amplio}
}