Direct answer
A human-in-the-loop agent proposes rather than acts. It assembles context and drafts a change with evidence attached; a named person approves or rejects with a reason; a separate executor carries out approved work. Most of the effort in these workflows is assembly rather than judgement, so the agent removes the slow part.
Human-in-the-loop AI agent
An agentic system in which a named human records an approval or rejection before any consequential action is executed. The agent holds no write tool; a separate execution layer performs approved work and records the resulting artifact.
Placing the gate so it stays useful
Approval applied everywhere trains reviewers to click through, which is worse than no gate because it looks like control. Place gates by consequence.
Gate it
Customer-visible changes, money movement, and anything difficult to reverse.
Usually do not
Reads, internal drafts, and reversible work with a cheap undo.
Decide by risk tier
Who may approve should scale with what the action can affect.
Attach the evidence
A reviewer deciding in moments needs the sources in front of them, not a summary to trust.
Why it does not slow the workflow down
The intuition that approval negates the benefit assumes the decision was the expensive part. Usually it is not: gathering context, checking records and drafting take the time, and a person deciding on a well-evidenced proposal is fast. Approval becomes a bottleneck only when applied to actions carrying no consequence.
The reviewer as the evaluation set
Rejections with reasons are the cheapest source of eval cases you will ever get, because each one is a real failure a domain expert already identified. Store the reason as structured data rather than free text where possible, and read the clusters.
The Propose–Decide–Execute pattern
CodeDTX builds agentic systems on the Propose–Decide–Execute pattern: agents may only write proposals with evidence attached, a named human records an approval or rejection with a reason, and a separate execution layer carries out approved work and logs the artifact. No agent holds a write tool to the outside world.
- 01
Propose
The agent analyses live system state and drafts a change, with the evidence it relied on attached to the proposal.
- 02
Decide
A named human approves, edits, or rejects with a reason. Risk tier determines who is allowed to decide.
- 03
Execute
A separate execution layer performs approved work — merge, publish, call, write — and records the resulting artifact.
- 04
Audit
Actor, reason, evidence, artifact, tokens, and cost are retained for every run, so any decision can be reconstructed later.
The six layers we build and review against
- 1
Agent layer
Agent architecture, tool use, memory and context, multi-agent patterns, structured outputs, orchestration.
- 2
Integration layer
MCP servers, tool contracts, API and database adapters, authentication, permissions, legacy system access.
- 3
Knowledge layer
Retrieval and RAG, vector and search architecture, enterprise knowledge sources, data access controls.
- 4
Reliability layer
Evals, tracing, observability, cost and latency budgets, fallbacks, regression tests.
- 5
Safety layer
Guardrails, prompt-injection defense, PII and data boundaries, human-in-the-loop gates, audit trails.
- 6
Product layer
The application people actually use: interfaces, approval queues, and operational runbooks.
Frequently asked
Does a human gate mean the agent is not autonomous?
It means autonomy is scoped. The agent decides how to assemble evidence and what to propose without supervision; it does not decide unilaterally to change a system of record. That is a deliberate boundary rather than an incomplete implementation.
Who should be the approver?
Whoever reviewed this work before the agent existed. They already hold the judgement and the accountability, and reusing them means the governance model does not have to be invented alongside the agent.
How do we stop reviewers rubber-stamping?
Gate fewer actions, attach real evidence to each proposal, and sample approved items in full rather than watching the approval rate. A high approval rate with thin review looks identical to a well-working system in a dashboard.
Have a workflow that should become AI-enabled?
Tell us about the system it lives in. We reply from an engineering seat, not a sales deck.