Objective-driven RSI
Iterate fast. Promote reliably.
Recursive self-improvement (RSI) means that agents improve the platform on their own in non-production. People choose which release candidates go to production. They promote them on their own schedule.
Objectives will steer the agents. Your organization will define them. An agent will read them through the Infrastream MCP or through Beck.
RSI does not run on its own today. The portal has no objectives yet. Agents do not merge on their own yet. The MCP runs only in the Infrastream integration environment. Each step on this page has a status tag.
Status tags
A status tag tells you what exists today. It does not tell you what we want. If the parts of a step have different statuses, the step shows more than one tag.
- LiveMerged. It runs, or a release includes it.
- Partly liveSome parts are live. Each part shows its own status.
- In integrationMerged. It rolls out to the integration environment.
- PlannedDesigned or listed. Not built.
The loop
The loop has seven steps. People act in steps 1 and 7. An agent opens the pull request in step 2. Machines act in steps 3 to 6. The arrow from step 7 to step 1 shows that the loop repeats.
Press Play to see a change move around the loop. Use Previous and Next to move one step at a time. If your system asks for reduced motion, the loop does not move and all steps are lit.
- Person
- Agent
- Machine
- A change
The loop has seven steps. Each step has a status tag. People act in steps 1 and 7. An agent opens the pull request in step 2. Machines act in steps 3 to 6.
The seven steps
The table gives the same facts as the figure. It names the actor and the status of each step.
| Step | Who acts | Status | What happens |
|---|---|---|---|
| 1. People set objectives | Person | Planned | People will set objectives, stories and acceptance criteria in the portal. An agent will be able to propose one. People will accept it.Objectives in the portal. The portal does not have them yet. |
| 2. An agent opens a pull request | Agent | Partly livePlanned | An agent opens a pull request under its own identity. In the plan, the agent also picks the next task.
|
| 3. CI and the image scan check the change | Machine | Live | CI and the image scan check the change. The scan uses a pinned Trivy. It fails closed: if it cannot run, the change does not pass.CI and the image scan run today. |
| 4. The change merges on green in non-production | Machine | Planned | An agent pull request will merge when the checks are green. This will apply in non-production only.The agent merge lane. It does not exist yet. |
| 5. The change goes to integration | Machine | Planned | Infrastream will apply the change to the integration environment, one apply at a time for each tenant.One apply at a time for each tenant. Today, applies that overlap can abort. |
| 6. The QA harness verifies the change | Machine | Planned | The QA harness will check the change in integration. It will use its own identity. It will publish evidence that is bound to the commit SHA.The harness exists, but it runs by hand today. The portal has the evidence rule. |
| 7. People promote release candidates | Person | In integrationPlanned | People promote release candidates to production, on their own schedule. In integration, Infrastream refuses a proven agent.
|
Parts of the loop by status
The list groups the parts of the loop by status. It shows what exists today and what we plan.
Live
- Step 2. The write_pr tool opens the pull request.
- Step 3. CI and the image scan run today.
In integration
- Step 2. The signed agent_id of the agent is in each caller assertion.
- Step 7. The portal refuses a proven agent that starts a release or acts on a stage after integration.
Planned
- Step 1. Objectives in the portal. The portal does not have them yet.
- Step 2. The agent does not pick the next task yet.
- Step 4. The agent merge lane. It does not exist yet.
- Step 5. One apply at a time for each tenant. Today, applies that overlap can abort.
- Step 6. The harness exists, but it runs by hand today. The portal has the evidence rule.
- Step 7. The release-candidate view.
People keep production
In integration, Infrastream refuses any proven agent that starts a release. It also refuses a proven agent that starts, approves, rejects, retries or rolls out a stage after integration. Only a person can do these actions.
A proven agent is a caller with a signed assertion, or an IAP service account. Every MCP client counts as an agent. For a human action, use the portal. See Agent access model for the rule and for the record that names the agent.
Objectives steer the agents
We plan that people set objectives, stories and acceptance criteria in the portal. An agent will be able to propose an objective. People will accept it.
The portal does not have objectives yet. When they exist, an agent will read them through the Infrastream MCP or through Beck.
Next
- Agent access model: what an agent can reach, and why production stays with people.
- The Agent Path: how your own AI agent will work through MCP, with a tier on each project.
- Agentic Governance
- CI/CD