← narwal.one/Second Brain
SecondBrain
Ask the Brain
Index/Conceptupdated Sun Sep 27 2026 08:00:00 GMT+0800 (Philippine Standard Time)

Human in the Loop

human-in-the-loopgovernanceautonomyagents

Human in the Loop

The pattern of keeping a human approval/review step inside an agentic workflow. Default operating model in 2026 enterprise AI per all three CXOTalk sources in this wiki.

When humans should stay in the loop (per CIO Agenda 2026 (CXOTalk))

  • High-risk, low-frequency decisions (Tim Crawford: "I don't think I want a fully autonomous surgeon anytime soon.")
  • Regulated functions (healthcare signoffs, legal review, finance approval)
  • Anywhere the cost of a single wrong action exceeds the speed benefit

When humans should be removed (same source, same speakers)

  • Time-critical functions where damage compounds while waiting (cybersecurity incident response)
  • Low-risk, high-frequency tasks where review doesn't add signal
  • Bounded use cases with reliable verification (see Bounded vs Unbounded Tasks)

The three-force model (Praveen, Agentic AI in the Enterprise (Praveen Akkiraju, CXOTalk))

  1. Workflow nature — bounded → less human; unbounded → more human
  2. Regulatory environment — even bounded tasks may require humans for accountability
  3. Harness quality — see Harness (LLM Agents); better harness → safer to remove humans

The Blitzy phased playbook (Autonomous Software Development with Blitzy (CXOTalk))

"You don't just flip a switch to full autonomy. That doesn't work that way. You have to build trust through a phased human-in-the-loop approach."

Stages:

  1. Trust-building: high-effort, low-risk friction (lang upgrades, doc gen, test gen)
  2. Role transition: developers shift from creators → editors → orchestrators
  3. New human role: prompt design + architecture review + execution validation

"Risk is too high for now" — the iterative go/no-go pattern

From Ben Mayrides (CISO, Cvent) in Governing AI Agents at Scale (Glean + Cvent, CXOTalk): H-I-T-L decisions should never be flat "no" — always time-bounded:

  • "Risk is too high for now" keeps the door open
  • Tech evolves, use cases evolve, delegation chains pop up
  • Go/no-go is iterative, not one-and-done
  • Cvent's working example: declined to deploy Anthropic Claude Co-work in regulated environments — for now

Operationally: pair every "not yet" decision with the conditions under which it would change.

The accountability case for HITL (Keen)

Five AI Risks That Can Get You Fired (IBM Technology) adds the missing accountability framing the other sources sidestep:

  • Unauthorized agentic AI is named as a top-five firing risk. The example: an agent "inadvertently sending out an email without a human in the loop to verify it first", or deleting a production file in a database
  • HITL isn't just risk-mitigation — it's the accountability seam. When the agent acts without a human, the question "whose name is on this?" gets harder to answer
  • The Zombie AI Agent failure mode is HITL gone to zero by neglect, not by design — and is firing-conversation territory

The Cherny/Boris view (next section) and Keen's view aren't in direct contradiction — Boris is arguing model improvements eventually eat the guard work; Keen is arguing the accountability role of HITL is independent of model quality. Worth distinguishing as a refinement.

HITL inside the coding loop — the SDD procedure (Everitt, 2026-09-26)

Full Course Spec-Driven Development with Coding Agents (DeepLearningAI) gives the most granular HITL procedure in the vault for software work, framed as "the agent generating something and us verifying it":

  • Human decides, agent interviews. Specs are drafted by the agent asking the human questions; the human reviews every spec file before implementation and commits it.
  • Permission prompts as control. Keep write/command approvals on unless you accept the security trade-off; "the ultimate responsibility for the code is yours."
  • Review at the intent level — does it match the spec — not variable names; "code you can commit under your name."
  • Small steps + clean breaks to keep review humane — the named human cost is Cognitive Debt / AI fatigue.
  • Agent-assisted verification — spawn subagents for a deep review "to validate that you weren't lied to"; the human then accepts or rejects their findings. A concrete instance of the agent-checks-agent pattern on Verification Tax.

Consistent with the Blitzy "creators → editors → orchestrators" arc above: Everitt casts the developer as architect/supervisor — "the agent is the muscle, but the spec is the brain."

2026-09-27 — "Human at the helm" critique (Goldman, How AI Is Changing Talent Not Just Tasks (HBR IdeaCast))

Paula Goldman (Salesforce) traces the phrase to the Cold War military (who makes the consequential call when a system detects an incoming missile?) and argues it degraded into "AI drafts, people approve" — which doesn't scale to agents doing many complex tasks at speed. Her replacement, Human at the Helm: route the right things to human judgment at the right time, plus deliberate friction on consequential decisions.

warning Tension with the per-output review patterns above Goldman's critique targets exactly the draft-and-approve loops prescribed by the SDD procedure and Keen's verify-before-send accountability case. Partial reconciliation: per-output approval survives where volume is low and stakes high; at fleet scale it becomes routing + exception review. See Human at the Helm.

Open question

Boris Cherny predicts the opposite trajectory (Boris Cherny on Coding Is Solved (Sequoia AI Ascent)): as models improve, all the human-in-loop scaffolding (prompt-injection guards, permission modes, review steps) becomes less important because the model just does the right thing. This is a directional disagreement worth tracking — see also the contradiction noted on Harness (LLM Agents).

Sources