Agent vs Human
Conformance Rate
The difference between agent DCS and human DCS — the single number that answers whether AI coding is making your architecture better or worse.
This page describes the intended design. It requires splitting CDA results by agent vs. human authorship, which isn't tracked yet. It doesn't appear on your dashboard yet, and no org currently has a live value for it.
− DCS(human-authored PRs)
What it signals
AHCR is the single most strategically important agentic metric rkito produces. It answers the question every CTO will eventually face: “Is AI coding making our architecture better or worse than humans?” A positive AHCR means agents (with MCP context) are outperforming humans on design conformance. A negative AHCR means they are not — and by exactly how much.
No other tool produces AHCR. It requires both agent attribution (knowing which PRs were agent-authored) and Lens-based CDA evaluation simultaneously. Code quality tools have neither. Architecture diagramming tools have neither. AHCR is rkito-exclusive.
The expected trajectory: AHCR starts negative (agents without design context produce more drift than humans who know the system implicitly). After rkito MCP integration, AHCR trends toward zero then positive (agents loaded with Ledger state, approved CHIs, and Lens criteria produce more conformant code than humans relying on implicit knowledge). Positive AHCR post-MCP integration is the ROI proof.
How rkito produces it
Every CDA audit is tagged by authorship. DCS is computed independently for agent-authored and human-authored PRs. AHCR is the delta — a single signed number that grows positive as MCP integration matures and agents load design context before the first token of code is generated.
Who this metric is for
AHCR is the number that answers the board question: "Are our AI tools making the codebase better or worse?" Not a sentiment, not a survey — a computed delta between agent and human design conformance. A positive AHCR is the quantified ROI of AI coding investment.
AHCR is the effectiveness signal for MCP integration. AHCR turning positive after MCP deployment proves the integration is working — agents are loading design context and producing more conformant code than humans relying on implicit knowledge. AHCR trending negative post-deployment means the MCP context is incomplete or misconfigured.
Watch
Do not use pre-MCP AHCR to restrict AI tools. The baseline negative AHCR is expected — it is agents without context versus humans with years of implicit knowledge. That is an unfair comparison. The valid comparison is post-MCP AHCR, where agents have equivalent access to design context. Making AI tool adoption decisions based on pre-MCP AHCR is like judging a new hire in their first week and firing them before they have been onboarded.
Also: AHCR can be negative because agents work on harder problems (complex boundaries with more Lens violations), not because they are inherently less conformant. Segment by work type before drawing conclusions.
What to do at each threshold
Step-by-step response playbook for each signal state.
Agents meeting or exceeding human baseline. If positive, quantify the margin as MCP ROI.
If AHCR is positive, express the margin as a concrete number for leadership: “agents produce conformant code X points more reliably than human-authored PRs.” This is the MCP ROI case — quantify it before the next budget cycle.
Monitor MCPCAR to confirm agents are actually loading design context. A positive AHCR with low MCPCAR is suspicious — either a measurement gap or agents are outperforming on a narrow slice of work.
Agents are meaningfully below human baseline. Find the pattern before diagnosing the cause.
Concentrated in specific Lenses: those rules need more explicit loading in agent MCP context. The agents have not been given the criteria they need before code generation starts.
Concentrated in specific agents: those agent configurations need MCP review. Distributed across all agents and Lenses: MCP integration may not be fully deployed.
Halt expansion of AI tool adoption scope until the gap is understood.
Check MCPCAR first. If agents are not loading MCP context, deploy MCP integration — the gap is structural absence of design context, not an agent capability problem.
If MCPCAR is high but AHCR is still deeply negative, check LSCR — agents may be loading Lenses but failing to produce conformant code despite having the criteria. That is a different problem requiring agent configuration review, not MCP deployment.
The companion agentic metric. ADRIFT measures ungoverned agent PRs as a percentage. AHCR measures conformance quality relative to humans. Together they give a full picture of agent governance health.
The leading indicator for AHCR. Low MCPCAR predicts negative AHCR — agents not loading design context before coding will produce less conformant code than humans with implicit knowledge.
The diagnostic tool when AHCR is negative. LSCR breaks down agent conformance by individual Lens rule, revealing which specific design criteria agents are failing to respect.
What is your org's AHCR?
rkito computes AHCR automatically by attributing every CDA audit to agent or human authorship and comparing the resulting DCS scores. Connect your repo and see the delta within minutes.