Executable
A governance judgment must affect the next step, not only produce prose.
Governance Engineering for Coding Agents
Every changemade by a Coding Agent,backed by inspectable evidence.
Weave goals, constraints, engineering methods, action judgments, and verification evidence into agent execution.
Judge before acting, change the next step during execution, and retain reviewable facts and unknowns at delivery.
Download for Windows
40-second real task
- Governance intentGoals and constraints
- Action judgmentFindings and methods
- Execution impactChange the next step
- Evidence recordFacts and unknowns
Governance Engineering
What is Governance Engineering?
Governance is not a report generated after the task. It participates before action, changes the next step during execution, and leaves reviewable facts and unknowns at delivery.
- 01Governance intent
Goals, constraints, acceptance
- 02Action judgment
Findings, methods, impact
- 03Execution impact
Change operations and checkpoints
- 04Evidence record
Execution facts, results, unknowns
Engineering means governance can enter execution, leave evidence, and withstand review.
Traceable
Move from a result back to judgments, tool calls, and source facts.
Reviewable
Separate confirmed facts, inferences, and unresolved unknowns.
Rules become part of how the Agent works instead of sitting outside the workflow.
A governance method must change the next step.
Finding: the change is complete, but there is no actual post-change test result.
Action effect: run the existing tests and use their output to decide whether the issue is fixed.
Change necessity
Avoid changing code for its own sake.
Systematic diagnosis
Make the next step answer the observed cause.
Structural entropy control
Limit unrelated changes and duplicate owners.
Evidence-backed delivery
Separate what ran from what it established.
Aegis × Autoloom
One governance approach. Two ways to use it.
Add a standalone method pack to an existing Coding Agent, or use Autoloom with governance checkpoints, trajectory, and records integrated into execution.
Aegis
Install governance methods in Coding Agent hosts that support Skills.
- Works across hosts
- Open methods and workflows
- Evolves in the Aegis repository
Autoloom
Combines methods, execution checkpoints, full trajectory, and governance records in a Windows product.
- Enters relevant judgment points during execution
- Keeps trajectory and governance records inspectable
- No separate Aegis installation
Both follow the same Governance Engineering direction while keeping product, installation, and release ownership clear.
One change leaves four layers of reviewable evidence.
These facts enter the Session record, connecting the result back to the original execution.
- 01
Goal and sources
Scope, acceptance, relevant sources
- 02
Action judgment
Finding, method, action effect
- 03
Execution facts
Tools, command results, file observations
- 04
Delivery record
Result, evidence references, unknowns
Trajectory makes execution reviewable.
Inspect context, model output, tool calls, arguments, results, and timing by turn.
Governance organizes scattered facts into one record.
Move from the overview into action assessments, checkpoints, execution facts, and work results.
Governance impact → Checkpoint → Execution facts → Result record
The record supports review; it is not a guarantee that the task is correct.
See Governance Engineering enter a real task in 40 seconds.
- 1Find the issue
- 2Form governance impact
- 3Apply the change
- 4Run existing checks
- 5Record the result
This demo3 failures → 4 passing1 implementation file changedExisting tests unchanged
The boundaries must be as clear as the governance.
- Permission mode
- Determines what the Agent may read, change, or execute.
- Model connection
- Uses a supported account or API configuration.
- Project location
- The workspace is local; model requests may send relevant context.
- Session record
- Tasks, tool results, and governance facts remain inspectable.
Start a reviewable task.
- 1Download the Windows client
- 2Connect a model
- 3Open a project and describe the goal and acceptance
The client is free. Your selected provider charges for model usage.
About Governance Engineering
Does governance slow the Agent down?
Governance enters only at relevant decision points and reuses the same Agent and Session. Actual cost depends on the task, model, and checks.
Does a governance record mean the result is correct?
No. It records grounds, execution facts, and unknowns. It does not replace user acceptance or guarantee correctness.
Where does project data go?
The workspace stays on your computer or connected SSH host. Model requests may send relevant context to the selected provider, depending on the task and configuration.
Which actions need authorization?
Permission mode determines what the Agent may read, change, or execute. Actions that need confirmation ask before execution.