Science ADK

Build, execute and audit autonomous scientific research.

An open source agent development kit that adds typed dataflow contracts, machine-recorded provenance and deterministic integrity gates to computational research — built directly on Google ADK.

Get started The 30-second example GitHub

---

A real result, in three commands

science-adk init my-kepler --example kepler
cd my-kepler
science-adk run

The run needs no API key, no account and no bundled dataset. It queries the NASA Exoplanet Archive for every confirmed planet with a measured period and semi-major axis, fits T = C·aᵅ in log-log space, and reports the exponent it recovers:

fetch:    fetched 3565 confirmed exoplanets from NASA
analyze:  fit 3565 planets: T ∝ a^1.4555, r² = 0.989997
evaluate: recovered exponent 1.4555 vs theoretical 1.5000, error 0.0445

The fit is blind — nothing in the pipeline assumes 3/2. Recovering Kepler's Third Law from measurements alone is the finding. Read the full walkthrough.

---

Why it exists

Coding agents can write analysis code quickly. Autonomous research needs something stronger: results that are computationally verifiable, reproducible, and defended against fabrication, leakage and silent numerical failure.

• Native Google ADK

  ---

  Every scientific agent is a real google.adk.agents.BaseAgent, running in the standard event stream and runner lifecycle, with no translation bridge.

  The foundation

• Deterministic integrity gates

  ---

  Six machine-checked conditions must pass before a number counts as a finding. Any failure scores the run 0.00, without exception.

  Gates and audit

• Provenance ledger

  ---

  Every input, fetch and tool call is recorded by the runtime, not by the agent. Fabricated results fail the tool_use_verified gate.

  Provenance

• Typed DAG validation

  ---

  Port types, cycles, dangling edges and unsafe source are caught before a single node executes.

  Workflows

• Git-native workspace

  ---

  Goal, hypotheses, DAG, runs, traces and learnings are plain files. Diff them, review them, reproduce them.

  The workspace

• Tools and MCP

  ---

  Every public function in tools/ is registered automatically; stdio and HTTP MCP servers plug in through science.toml.

  Tools

---

The shape of an experiment

Research is expressed as a directed acyclic graph of specialised agent primitives, each with typed input and output ports:

graph LR
    Config[ConfigAgent<br/><i>parameters</i>] --> Data[DataAgent<br/><i>ingestion / simulation</i>]
    Data --> Algorithm[AlgorithmAgent<br/><i>method under test</i>]
    Algorithm --> Evaluate[EvaluationAgent<br/><i>metric calculation</i>]
    Evaluate --> Visualize[VisualizationAgent<br/><i>figure rendering</i>]

This is the analyze node of the bundled example, near enough verbatim:

from science_adk import AlgorithmAgent, Port, Ports


class FitKeplerLaw(AlgorithmAgent):
    """Recover the power-law exponent from a blind log-log fit."""

    ports: Ports = Ports(
        inputs=[Port("planets", "json")],
        outputs=[Port("fit_result", "json", "Power-law exponent, r², diagnostics")],
    )

    async def execute(self):
        planets = await self.input("planets")
        periods = [p["period_days"] for p in planets]
        axes = [p["semi_major_axis_au"] for p in planets]

        fit = await self.call_tool("power_law_fit", x=axes, y=periods)

        await self.log(f"fit {fit['n']} planets: T ∝ a^{fit['exponent']:.4f}")
        return {"fit_result": fit}

---

Where to go next

• New here

  ---

  Install, run the Kepler example, then scaffold your own project.

  Get started

• Want the ideas first

  ---

  Ports, provenance, gates, the audit, and why each one exists.

  Concepts

• Building something

  ---

  Task-shaped guides: write agents and tools, validate, run, audit, evolve.

  Guides

• Using a coding agent

  ---

  Install the seven research skills into Claude Code, Cline, Cursor, OpenCode or Antigravity.

  Skills