Skip to main content
infy governance is not only for infy agents. infy.integrations wraps the tool or action chokepoint of frameworks you already use, so you can add deny-by-default policy, human approval, and a tamper-evident audit to agents you already run, without changing them.

​
The pattern

Every governed framework follows the same shape. You build a Governance object once, then pass your existing tools through a govern adapter. The adapter returns drop-in replacements that route each execution through governance first.
A denied or unapproved action never runs. The model receives the block reason as the tool output and adapts. Every decision, allowed or denied, lands in the same SHA-256 or HMAC hash-chained audit you can verify().
The core stays zero-dependency. Each adapter lazily imports its framework, so nothing is pulled in until you import the adapter you need (for example infy.integrations.smolagents).

​
How govern works

govern takes your list of tools and the Governance object, plus an optional risk map, and returns wrapped tools.
1

You pass tools plus a risk profile

risk maps a tool name to {"verb": ..., "risk_tier": ..., "side_effect": ...}. This is the static metadata infy’s risk engine reads to decide whether a call is high enough risk to need a human.
2

Unprofiled tools fail safe

Any tool not named in risk gets the default profile, which is conservatively high-risk and side-effecting (verb: "EXECUTE", risk_tier: "high", side_effect: True). An unprofiled tool is escalated to a human approver rather than waved through.
3

Execution passes through the chokepoint

Each wrapped tool calls governance.before_tool(...) first. If the decision is not allowed, the original tool never runs and the block reason is returned to the model. If it is allowed, the tool runs and governance.after_tool(...) records the result.
You can also override the fallback profile with the default argument:
An unprofiled side-effecting tool is treated as HIGH risk on purpose. If your default approver is DenyAll, an unprofiled tool is denied. Profile every tool you intend to allow, or set an approver that can grant a human yes.

​
What you get

The same governance guarantees that apply to native infy agents apply to wrapped ones:
  • Deny-by-default, fail-closed. Any error in policy, risk, or approval yields a deny, and is still audited. There is no path to a silent allow.
  • Human approval gate. A pluggable Approver pauses high-risk calls for a human yes or no, recorded in the audit chain.
  • Tamper-evident audit. Every decision is appended to a hash-chained log you can verify().
The wrapper adds about 50 microseconds per tool call, so it is cheap enough to leave on.

​
Choose your framework

smolagents

Wrap smolagents tools with infy.integrations.smolagents.govern. Governance gates each tool at Tool.forward.

LangChain

Wrap LangChain tools with infy.integrations.langchain.govern, the same pattern at the LangChain tool boundary.

OpenHands

Plug infy governance into OpenHands through infy.integrations.openhands.build_analyzer, an adapter for its SecurityAnalyzer hook.
Runnable demos, including a real smolagents agent that has a destructive action denied and a deploy paused for a human, are in the repository examples/ directory.