ActionGraph
Every decision remembered. Every action earned.
The open-source memory and conscience for AI agents.
Intelligence is cheap now. Trust isn't.
For the first time, software can decide. It can issue the refund, roll back the deploy, grant the access, reroute the truck. The models are ready. The world isn't, because nobody can answer three simple questions about an agent:
- Should it be allowed to do this?
- Why did it do that?
- Has it earned the right to do it alone?
Today the answer is a toggle someone flipped and a log nobody reads. That's why most agents still wait for a human to click "OK" on every action, and why the ones that don't keep their builders up at night.
ActionGraph answers those three questions with evidence. Before an agent acts, it asks. It gets the options your rules allow, ranked by what actually worked in cases like this. It records what it chose and why. A person approves where it matters. The outcome comes back and is graded. And autonomy, the right to act alone, is earned one proven decision at a time, and taken back the moment results slip.
ask ──▶ record why ──▶ person approves ──▶ outcome ──▶ earned autonomy
▲ │
└──────────────────────── every decision makes the next one smarter ┘
Imagine a world that runs on ActionGraph
Every hospital, bank, airline and factory runs on decisions: millions a day, each one made, mostly forgotten, rarely measured. Now imagine every one of them, whether made by a person or an agent, landing in an open, shared record of what was decided, why, by whom, and whether it worked.
- Companies stop forgetting. When your best claims adjuster retires, their judgement doesn't leave with them. Ten thousand of their decisions, with outcomes, train the next adjuster and the next agent.
- Trust becomes measurable. "Can this agent approve refunds under $200?" stops being a meeting and becomes a number: worked 412 of 431 times, 90% confident it clears the bar.
- Autonomy grows safely, like a career. Every agent starts as an intern. It proposes, a person approves. As its record proves itself, it earns more room, one decision type at a time. A bad week costs it that room automatically.
- Accountability is built in, not bolted on. An auditor asking "why did the AI do this?" gets the full answer in one read: the situation, the options, the rules checked, who approved, and what happened next.
- Humans move up the stack. People stop rubber-stamping the routine and spend their attention where it counts: the new, the risky, the irreversible.
That's enterprise autonomy: not AI that replaces judgement, but AI that inherits it, proves itself, and earns its freedom. ActionGraph is the open layer that makes it possible, the way TCP/IP made the internet possible: boring, reliable, everywhere, and owned by no one.
Three principles we won't bend
- No agent grades its own homework. Outcomes come from the systems that saw them. An agent's own reports earn it nothing, and hidden random audits check its word.
- No agent approves itself. Ever. Approvals bind to the exact request a person saw.
- Autonomy is a record, not a setting. It is granted by evidence with stated confidence, and revoked by evidence just as fast.
See it in 30 seconds
pip install actiongraph-ai
from actiongraph import ActionGraph, ApprovalRequired
DB = "sqlite:///quickstart.db" # a local file: no server, no API key
agent = ActionGraph(DB, agent="support_bot") # it can ask, never approve
person = ActionGraph(DB) # a person: their name goes on the record
@agent.guard(subject="Order:{order_id}", autonomy_after=3)
def issue_refund(order_id: str, amount: float):
print(f" refunded ${amount} on {order_id}")
for order_id in ["A-1001", "A-1002", "A-1003", "A-1004"]:
try:
issue_refund(order_id, 40.0)
print(f"{order_id}: ran on its own: the agent earned it")
except ApprovalRequired as e:
print(f"{order_id}: waiting for a person")
person.approve(e.decision_id, by="dana")
issue_refund(order_id, 40.0)
A-1001: waiting for a person
A-1002: waiting for a person
A-1003: waiting for a person
A-1004: ran on its own: the agent earned it
Three approved runs that worked, and on the fourth the agent acts alone. One failure, and it asks again. That's the whole idea, in fifteen lines.
What you get
| A decision ledger | Every decision kept whole: the situation, the options offered, every rule check, who decided, who approved, why, and the outcome. Exportable as open decision events. |
| Evidence, not opinion | Options ranked by how often they worked in cases like this, with honest confidence. When the record is thin, it says so. |
| Playbooks | Versioned YAML: the decisions you govern, the options allowed, rules that block or route to a person, and what "worked" means. A person activates each version. |
| Earned autonomy | recommend → approve → auto, per agent and per action, earned from outcomes with 90% confidence and taken back automatically. |
| Learned ranking | Optional contextual RL that learns from outcomes, can start from an outside model like Jev, and ranks only after a person promotes it. |
| Flows and swarms | Multi-step, multi-agent processes with gates. Agents vote; agreement decides, a split goes to a person. |
| From approval to execution | Every answer says what happens next. A decision hands the executor a single-use grant for exactly the approved action and parameters, and a webhook tells it to run. |
| Works with every agent | Claude, GPT, LangGraph, CrewAI or your own code: in-process Python, HTTP, or MCP. actiongraph mcp-gate puts a person in front of any MCP server in one line. |
Explore in nine examples
Each runs on its own with nothing but pip install actiongraph-ai.
| Example | Shows |
|---|---|
| 1. Your first approval | A guard, approve and reject, the record |
| 2. Rules and evidence | Options ranked by what worked; a rule that blocks; a rule that routes to a person |
| 3. Earning autonomy, and losing it | recommend → approve → auto, and auto taken back when results turn |
| 4. Who grades the agent? | Its own word earns nothing; hidden audits; the honesty test |
| 5. Safety rails | No self-approval, approvals bound to the request, one approval, one run |
| 6. Learned ranking | Contextual RL, with and without a Jev prior |
| 7. The decision ledger | One entry in full, its timeline, totals, export and ingest |
| 8. From approval to execution | next on every answer, and a single-use grant to run exactly what was approved |
| 9. Approvals in Slack | Approve and Reject in Slack, then a webhook hands your executor the grant |
Where it's going
ActionGraph today is the decision record and authority layer for one organization. The road ahead:
- Tamper-evident ledger: hash-chained decision events, so the record proves it hasn't been edited.
- Decision intelligence at scale: discovery of the decisions your company already makes, and backtesting against your own history.
- A hosted version: many organizations, per-person sign-in, Slack approvals, on-call routing and outcome connectors.
- An open standard: decision events any agent framework can emit, so every decision, everywhere, can be remembered and learned from.
Honest about today
ActionGraph is a beta (0.2), a research prototype. Use it today with a person approving, and as the record of what your agents decided. It holds against an agent when run as a server with separate keys; on one machine with no server it is a record, not a lock. The ledger isn't tamper-evident yet. Read What it protects, and what it doesn't before letting an agent act alone on anything costly or irreversible. Provided as is, without warranty (Apache-2.0).
Go deeper
- The full technical reference: docs/REFERENCE.md: guards, MCP gate, playbooks, flows, server, dashboard, API
- Docs · Playbook spec · Decision events spec
- Fourteen scenarios · How good is the evidence?
Join us
The future where agents earn trust instead of being granted it gets built in the open. Star the repo, try an example, open an issue, send a pull request (CONTRIBUTING.md). Report security problems privately (SECURITY.md).
Software learned to think. ActionGraph teaches it to be trusted.
Metadata
Release files for actiongraph-ai 0.3.0
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | Actions |
|---|---|---|---|---|
| actiongraph_ai-0.3.0.tar.gz | 486.1 kB | With provenance | Details for actiongraph_ai-0.3.0.tar.gz |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| actiongraph_ai-0.3.0-py3-none-any.whl | Python 3 | none | any | Details for actiongraph_ai-0.3.0-py3-none-any.whl |
Total release size: 926.6 kB
Release files / actiongraph_ai-0.3.0.tar.gz
| Download URL | actiongraph_ai-0.3.0.tar.gz |
|---|---|
| Size | 486.1 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
7ee1ce83becb8518378b91d750e09942ae1adde477a5884f6350b2a4d36965fe
|
|
BLAKE2b-256 checksum How to use checksums |
a75d2d0bf3ba6c5a47848bb14abd40cd29c079bca9b2c8bb05ed37b1225e6e5b
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.16
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Nov 15, 2026.
Transparency logRelease files / actiongraph_ai-0.3.0-py3-none-any.whl
| Download URL | actiongraph_ai-0.3.0-py3-none-any.whl |
|---|---|
| Size | 440.5 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
4acf02cdc2a1152304e6f7dec887e2bcd9055544e1d77b9cb621c6b91b45c08e
|
|
BLAKE2b-256 checksum How to use checksums |
5a6735639e9ac07a8d7e026a8f79cd327b2d62b0ebbb2ca3a094e27321b8b576
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.16
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Nov 15, 2026.
Transparency logSecurity for actiongraph-ai 0.3.0
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
Full provenance
Both files in this release have been uploaded with provenance metadata.
Attestation for actiongraph_ai-0.3.0.tar.gz
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Nov 15, 2026.
Transparency logA note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.