HHI™ Applied

Claude Code chose deliberate over sophisticated, and scored closer to meaningful oversight for it

Claude Code · June 2026 · HITL Maturity Model™ · HHI™ v1.0 · Intellectus Advisors

HHI™

50Managed

Dimension Breakdown

  • D1Intervention Quality62
  • D2Feedback Loop Integrity38
  • D3Human Capacity Health55
  • D4Trust Calibration58
  • D5Governance Compliance40
  • D6Agentic Readiness50

Claude Code is assessed as a single archetype. Unlike Codex, the product does not offer differentiated compliance infrastructure between individual and enterprise tiers. The per-command approval model, CLAUDE.md governance, and git-level authorship are available identically to solo developers and enterprise teams. The HHI™ score of 50 reflects the product's oversight posture across both user types.

Claude Code made a different architectural bet than Codex. Instead of optimising for agent autonomy with auto-review as a safety net, Anthropic shipped a tool that asks the developer to approve each tool call: file edits, shell commands, network requests, before they execute.

The result is a less impressive demo and a stronger oversight profile. HHI™ 50, Managed. The human is genuinely in the loop. The gaps lie elsewhere.

Claude Code's per-command approval is the simplest possible oversight mechanism. It is also one of the few that actually works.

Where the design pays off

D1 (Intervention Quality) and D4 (Trust Calibration) are the strongest dimensions in the assessment. Developers report that the approval prompt forces a meaningful read of the proposed action, not a glance, a read. Trust is calibrated against observed behaviour rather than marketing claims.

Where it falls short

D2 (Feedback Loop Integrity) and D5 (Governance Compliance) drag the composite down. Approvals and denials are not retained as structured evidence. There is no standard export for audit. Reviewer dissent disappears the moment the next prompt arrives. For regulated workflows, this is a blocker.

Compared with Codex

Claude Code beats Codex on the dimensions that matter most for human oversight (D1, D2, D4) and trails on the agentic-architecture dimension (D6). The trade-off is honest. Buyers choosing between the two should decide whether they are buying an agent or a copilot.

OBSERVE Findings

Seven dimensions of human oversight.

  • O

    Output

    Every consequential output passes through a human approval prompt before execution.

    Per-command approval is structural, but it is not frictionless in the right direction for all developers. Senior engineers who understand the codebase deeply engage with each approval meaningfully. Developers working in unfamiliar codebases may approve commands without fully understanding their consequences, not from fatigue but from knowledge gaps. The oversight mechanism is only as strong as the reviewer's ability to evaluate what they are approving.

  • B

    Burnout

    Approval fatigue exists but is bounded; developers report higher sustained attention than with auto-review tools.

    The terminal-first workflow creates more natural pause points than cloud-parallel models, but session-level permission grants can erode those pause points progressively. A developer who starts by approving each shell command individually, grants session-level permission after the third session, and then grants permanent permission after the tenth has moved from deliberate oversight to structural bypass without noticing. No product signal marks this transition.

  • S

    Signal

    Tool-call telemetry is rich but not exposed to the user as oversight signal.

    Git-level AI authorship attribution is the strongest per-decision provenance signal of any assessed coding tool. But provenance is not a feedback loop. Knowing that Claude authored a particular commit does not tell the team whether that commit introduced a logic error that propagated into production, or whether the pattern Claude used in that commit has since been identified as problematic. Provenance answers "who did this?" Feedback loops answer "what did we learn from it?" Claude Code has the former and lacks the latter.

  • E

    Error

    Caught errors are visible in-session; cross-session learning loops are absent.

    CLAUDE.md serves as a persistent correction layer, but its effectiveness depends entirely on how it is maintained. Teams that treat CLAUDE.md as a living governance document, reviewing it quarterly and encoding corrections systematically, will see D2 scores materially higher than teams that update it reactively. The product supports both uses. The score difference is in the practice, not the feature.

  • R

    Regulatory

    No native audit export. Compliance teams must build their own evidence pipeline.

    For teams preparing for EU AI Act compliance reviews, the git-level audit trail has a meaningful advantage over centralised log exports: it is durable, distributed, and not subject to a 30-day retention window. Every AI-authored commit is permanently attributable. The limitation is aggregation. A compliance reviewer cannot run a query across git history the way they can run a SIEM query. The investment needed is a git analytics layer, not a new logging infrastructure.

  • V

    Variation

    Approval threshold is uniform across action types; no risk-weighted gating.

    Claude's in-language uncertainty expression, including hedging, flagging unknowns, and declining to proceed when context is insufficient, is a genuine D4 signal, but it is not a calibration mechanism. A developer who has worked with Claude Code for six months has learned Claude's hedging patterns and may be filtering them out. The absence of a formal uncertainty score means there is no external reference point. The developer cannot know whether Claude's expressed confidence today is calibrated against Claude's historical accuracy in similar situations.

  • E

    Evolution

    Roadmap signals continued bias toward explicit human control, not more autonomy.

    The terminal-first workflow creates more developer-to-implementation contact than cloud-parallel delegation models. Developers see the code being written, the commands being proposed, the errors being encountered. This contact is not the same as expertise. Watching Claude debug a problem is not the same as debugging it yourself. But it is closer to expertise maintenance than post-hoc PR review. The D3 risk for Claude Code is lower than for parallel-task delegation tools, but it is not zero and it is not monitored.

Why deliberate is underrated

The conversation about agentic coding tools in 2026 is dominated by capability metrics: tasks per hour, lines of code generated, PR throughput. Claude Code scores poorly on all of these relative to Codex. It does not run parallel tasks. It does not operate autonomously while the developer has coffee. It does not queue work overnight. It sits in the terminal and waits to be asked.

From a throughput perspective, this is a limitation. From a human oversight perspective, it is a design choice the HHI™ framework rewards. The per-command approval model is not friction. It is the human oversight mechanism working as intended. Every consequential action requires a deliberate yes from a human who understands what they are approving. That is the closest any current agentic coding tool comes to what EU AI Act Article 14 actually envisions.

The question is not whether Claude Code is more capable than Codex. It is whether the oversight posture it creates is more defensible, and whether that defensibility has value in the regulatory environment that enterprise AI teams are now operating in.

The CLAUDE.md governance opportunity

CLAUDE.md is Claude Code's primary mechanism for persistent team instructions. At most organisations it functions as a productivity tool, encoding preferences, style guides, and recurring patterns so developers do not repeat themselves. But CLAUDE.md has a more significant governance use that is largely untapped: it is the only native location in Claude Code where an organisation can encode its human oversight policy.

A CLAUDE.md that includes explicit escalation thresholds (always ask before modifying authentication logic), documented permission boundaries (never write to production databases without explicit confirmation), and correction records (previous attempts at X produced Y issue, require human review for this pattern) is doing governance work, not just productivity work. It is creating a documented oversight mechanism per use case, which is exactly what Article 14 and Article 26(2) require.

Most teams are not using CLAUDE.md this way. The gap is not a product gap. It is a governance awareness gap. And it is closeable with a half-day of deliberate work by anyone who has read the EU AI Act's human oversight obligations.

What the EU AI Act requires, and where Claude Code stands

EU AI Act Article 14 requires that high-risk AI systems include human oversight measures that enable natural persons to monitor, intervene, and override the system's outputs. Article 26(2) requires deployers to implement those measures. Article 26(6) requires six months of log retention.

Claude Code's per-command approval model satisfies the spirit of Article 14 more directly than any guardian-agent architecture currently on the market. A human natural person is making a deliberate decision before each consequential action. That is not a proxy for oversight. It is oversight.

The Article 26(6) log retention gap is real and requires deployer action. Git history provides timestamped, attributable records of every AI-authored change, but git is not a centralised audit log system. Enterprise teams need to instrument a log aggregation layer that captures session-level activity, command approvals, and CLAUDE.md governance decisions in an exportable, SIEM-compatible format. This is deployer-built infrastructure, not a platform feature.

The honest assessment: Claude Code is closer to Article 14 compliance by product design than any alternative currently assessed. It is further from Article 26(6) compliance by infrastructure than Codex Enterprise. The right choice for a regulated organisation depends on which article they are more exposed on, and which gap is cheaper to close.

The D6 ceiling and why it matters for 2027

Claude Code's D6 (Agentic Readiness) score of 50 reflects a product that handles current agentic complexity well: sequential task execution, per-command approval, explicit permission management. It is not yet designed for the multi-agent, long-horizon workflows that are already entering enterprise production environments.

The gap is specific. In a Claude Code session, the developer sees what Claude proposes and approves each action. In a multi-agent workflow where Claude Code is one agent among several, receiving instructions from an orchestrator and passing outputs to downstream agents, the per-command approval model breaks down. There is no mechanism for the human to see the orchestrator's plan before execution begins. There is no visibility into sub-agent activity. There is no checkpoint architecture for long-horizon tasks that span hours or days.

This is not a current failure. Most Claude Code users are not running multi-agent workflows today. It is a forward-looking gap. The NIST AI 600-1 agentic controls (AGT-021, AGT-012, AGT-017) provide the framework for what this architecture needs to look like. Teams selecting Claude Code for agentic work that will scale should be asking Anthropic's roadmap questions about multi-agent oversight before they are locked into a workflow that outgrows the product's current oversight model.

What moves the score from 50 to 65

Three changes would materially lift Claude Code's HHI™ from 50 toward 65 and above. None of them require a fundamental product change.

First, a centralised log aggregation layer. Git history is necessary but not sufficient for enterprise governance. A lightweight integration that captures session-level approval events, CLAUDE.md change history, and command execution records into an exportable JSON format would close the D5 gap. This is a deployer-built component, not a product request, and it can be built in a week by a team that knows what they need.

Second, a systematic CLAUDE.md correction pipeline. The current pattern has developers updating CLAUDE.md reactively when a recurring error becomes frustrating. A practice of weekly CLAUDE.md review, where teams audit recent correction patterns and encode them proactively, would close the D2 gap and create a self-improving oversight layer. The product supports this. Most teams are not doing it.

Third, an explicit trust calibration practice. Session-level permission grants, where developers batch-approve categories of commands, are the primary trust-drift mechanism in Claude Code deployments. A quarterly audit of granted permissions, with explicit decisions about which should be revoked or narrowed, is governance hygiene that takes two hours and materially reduces the D4 risk. No product feature required.

Key Findings

Per-command approval works

The simplest oversight mechanism in the market is also the most effective one we tested.

Feedback loop is the weak link

Without evidence retention, oversight events generate no organisational learning.

Compliance gap is fixable

An audit export and structured dissent capture would lift the score into Healthy territory.

Honest architectural trade-off

Claude Code is slower and less autonomous. It is also closer to genuinely supervised AI engineering.

References and further reading

  • [1]Anthropic. (2026). Claude Code documentation, permissions and trust model. docs.anthropic.com/claude-code
  • [2]Anthropic. (2026). Claude Code, system prompt and CLAUDE.md governance. docs.anthropic.com
  • [3]Belagodu, S. (2026). HITL Maturity Model™ v2 and HITL Health Index™, framework documentation. Intellectus Advisors. humynpulse.com
  • [4]Augment Code. (2026, June). 7 AI coding tools for EU AI Act compliance. augmentcode.com
  • [5]EU AI Act. (2024). Article 14, Human oversight. EUR-Lex. eur-lex.europa.eu
  • [6]NIST. (2023). AI Risk Management Framework. nist.gov/artificial-intelligence
SB

Su Belagodu

Founder · Intellectus Advisors

Apply It To Your AI

Run the same assessment on your own AI initiative.