Formula
AIFI = (agent-authored PRs where
CHI approval date
< agent first commit date
÷ total agent-authored PRs)
× 100Thresholds
Agents executing within approved scope
Orchestration gaps — investigate workflows
Autonomous generation without scope — highest risk
Reading it
The agentic equivalent of IFR. AIFI measures whether AI agents are being invoked against architect-approved Change Intentions before generating code — not after. A high AIFI means the orchestration layer ensures an approved CHI exists before the agent starts. A low AIFI means agents are autonomous code generators operating without pre-approved architectural scope: the highest-risk agentic deployment pattern.
IFR and AIFI look similar but require different remediation. Low human IFR is a culture and process problem — developers are not following the intent-first workflow. Low AIFI is an orchestration tooling problem — the system that spawns agents is not checking for approved CHI scope before dispatch. The fix is not developer education; it is workflow integration: require get_approved_chi() as a pre-start check in the orchestration layer.
AIFI rising after rkito MCP integration is evidence that the integration is working as intended. Agents invoked with an approved CHI in scope know what they are authorised to change before writing a single line. They operate as specification executors, not autonomous generators. This is the distinction that separates governed agentic development from ungoverned agentic development.
How rkito produces it
Agent-authored PR attribution uses commit signatures and MCP access logs to identify agent-generated branches. CHI approval timestamps are then compared against the earliest agent commit on that branch. AIFI is the percentage of agent-authored structural PRs that are AIFI HITs, computed per sprint.
- 01
Agent-authored PR identified
Commit signatures and MCP access logs confirm agent authorship of the branch
- 02
CHI approval timestamp retrieved for the branch
rkito resolves the Change Intention linked to the branch and reads its architect approval timestamp
- 03
Agent first commit timestamp retrieved for the branch
Earliest agent commit on the branch is recorded as the agent start time
- 04HIT
CHI approval date < first commit date
AIFI HIT — agent invoked within approved architectural scope
- 05MISS
CHI approved after first commit, or no CHI
AIFI MISS — agent operated as autonomous generator without pre-approved scope
- 06
AIFI = % of agent-authored structural PRs that are AIFI HITs, per sprint
Hit rate reported per sprint and trended on the org dashboard
Who reads it
AIFI shows whether the team uses agents as specification executors (high AIFI) or autonomous generators (low AIFI). These are fundamentally different risk profiles.
AIFI is the signal for whether agent sessions start with approved design context. Low AIFI = agents are making architectural decisions without review.
How it misleads
What to do at each threshold
Step-by-step response playbook for each signal state.
Agents are being invoked within approved architectural scope. Confirm MCPCAR is also high — agents should be loading design context, not just having a CHI on record.
- Confirm MCPCAR is also high — agents should be loading design context, not just having a CHI technically on record. A high AIFI with low MCPCAR means the pre-start check exists but the context load does not. The CHI is approved; the agent never read it.
Identify which agent workflows are starting without approved CHIs. Integrate the pre-start check in the orchestration layer.
- Identify which agent workflows are starting without approved CHIs. Integrate
get_approved_chi()as a mandatory pre-start check in the orchestration layer.
Agents are operating as autonomous generators without design scope. Require the orchestration layer to call draft_chi and halt if no approved CHI exists before agent startup.
- Require the orchestration layer to call
draft_chiand halt if no approved CHI exists before agent startup. This is an orchestration tooling problem, not a culture problem — the fix is a code change, not an announcement.