Staying in the Loop with Agents

When an agent does the work, the human's job shifts from writing to understanding. If you skip that, you become a rubber stamp who slowly loses the thread of your own codebase. The fix is to make the agent teach the work back to you and verify you actually got it — before the session ends.

The failure mode

As agents take over implementation, the default human posture degrades to approving diffs you don't fully understand. Agent Operations Skills names the endgame: "the human becomes the memory layer, which defeats the point of having an agent" — except here it's worse, because the human stops being even that. You lose the ability to debug, extend, or make good decisions about a system you nominally own. Velocity without understanding is debt. Source: compiled from Agent Operations Skills, 2026-06-01

The technique: make the agent teach you

The strongest countermeasure is to flip the agent into a teacher whose goal is your comprehension, not more output. The pattern combines three well-studied learning mechanisms:

Anthropic's Thariq (@trq212) surfaced an internal example from Suzanne: a prompt that makes Claude teach the work back so the human stays in the loop with what was shipped. Thariq notes Suzanne pairs it with voice mode for more natural Q&A. Gist linked from thread (17,809 bookmarks on the bookmark tweet). Source: X/@trq212 thread, 2026-06-01

• Active recall — you restate your understanding and answer quiz questions before being told the answer. Retrieval, not re-reading, is what builds durable understanding.
• The Feynman technique — explaining a concept in plain terms (eli5 / eli14 / elii) exposes exactly where your model is fuzzy.
• A completion gate — the session cannot end until mastery is demonstrated, so "I think I get it" never substitutes for evidence (mirrors verification-before-completion, applied to the human).

What "understanding" must cover

Both levels, never just one:

• High level — motivation: why this problem mattered, what it unblocks.
• Low level — business logic, edge cases, the actual code paths.

Organized into three buckets: (1) the problem — what it was, why it existed, the branches considered; (2) the solution — what changed, why resolved that way, the design decisions and trade-offs, the edge cases; (3) the broader context — why it matters and what it will impact (blast radius). Prioritize the problem: a human who owns the problem can re-derive the solution. Source: User favorite prompt, 2026-06-01

Participation Artifacts

Geoffrey Litt's July thread widens this page from "quiz me after the work" to "make the work participable while it is happening." If agents write most of the code, the scarce human contribution becomes judgment over intent, architecture, taste, and risk. That cannot be recovered from a raw diff alone. The agent needs to produce artifacts that help the human re-enter the problem space quickly enough to shape the next move. Source: X/@geoffreylitt, 2026-07-02

The thread's design direction is a menu of legibility artifacts:

• Explain-diff and literate diffs: show the narrative of the change, not only changed lines.
• Background before detail: teach the intuition, constraints, and alternatives before asking the human to judge implementation.
• Quiz as speed regulator: make comprehension an explicit gate, but use it to tune pacing rather than punish the reviewer.
• Micro-worlds and interactive figures: let humans manipulate the decision space when static prose is too weak.
• Shared workspaces: agents should work in a surface like Notion, Linear, or a wiki where intermediate artifacts can be inspected, edited, and remembered.

This complements Agent Legibility: legibility is the harness property; participation artifacts are the human-facing output form. The local rule is to request these artifacts at irreversible boundaries: API shape, data model, migration plan, security posture, or UX direction. Source: X bookmark artifact audit, 2026-07-06

Operating principles

• Incremental, not end-loaded — confirm mastery of each stage before advancing; don't dump everything at the end.
• Restate-first — have the human state their understanding before you explain, to find the real gaps.
• Drill the whys — chain "why" down to root motivation, then cover what and how.
• Quiz honestly — open-ended and multiple choice; vary the correct-answer position; hide answers until submitted.
• Show evidence — open the code or the debugger when a concept is load-bearing.

Where this lives

The executable owner is the Security and Review Skills skill (/teacher, "teach me this session", "make sure I understand"). The compact product route is Security and Review Skills. Do not create a separate learn-quiz skill: the gist is the upstream source for this same pattern, and MECE skill ownership keeps it folded into the existing teacher/understanding route. It composes with Plan Mode Review Checklist (review-time) and The Brain-Agent Loop (the wiki captures what was learned so the understanding compounds across sessions instead of evaporating).

For live approval, interruption, and cancellation during a run, use Human-in-the-Loop Control. This page is about maintaining human comprehension; HITL control is about where the human has authority or accountability before the agent acts. Agent-Ping is the tiny cross-harness CLI primitive that makes that control reachable from any shell-capable agent.

Why it matters here

Kevin's whole system is built so he "should never have to remember anything" operationally — but that is the opposite of not understanding. The wiki offloads bookkeeping; teacher-mode protects comprehension. The two together let him move fast with agents without going blind to what they ship. Source: compiled from Agent Soul + Agent Ethos, 2026-06-01

Concept Position

---

Timeline

• 2026-07-06 | Absorbed Geoffrey Litt's participation-artifacts sequence as one thread. Updated the page from end-of-run teacher-mode alone to a broader loop: agents should produce explain-diff, background, quiz, interactive, and shared-workspace artifacts so the human can still participate in problem framing and judgment. Source: X/@geoffreylitt, 2026-07-02
• 2026-07-01 | Concepts category refresh added this page to the Brain, memory, and retrieval family, linked it to Concept System Map, and kept it standalone because it owns this reusable mental model: When an agent does the work, the human's job shifts from writing to understanding. If you skip that, you become a rubber stamp who slowly... Source: User request, 2026-07-01
• 2026-06-30 | Deep-reviewed the Learn Quiz gist and screenshot. Decision: update existing Security and Review Skills / Security and Review Skills routes rather than add a duplicate learn-quiz skill. Source: X bookmark artifact audit, 2026-06-30
• 2026-06-01 | Page created from a favorite "teacher" prompt Kevin shared. Compiled the failure mode ("rubber stamp / memory layer"), the learning science (active recall + Feynman + completion gate), the two levels and three buckets of understanding, and the link to the Security and Review Skills skill. Source: User, 2026-06-01
• 2026-06-01 | @trq212 shared Suzanne's Anthropic-internal "stay in the loop" prompt; 10,413 likes, 17,809 bookmarks. Voice-mode pairing noted. Source: X/@trq212, 2026-06-01
• 2026-06-12 | Top-25 bookmark absorption cross-linked from ai-agents hub (#20). Source: x-bookmark-absorb, 2026-06-12
• 2026-06-17 | Referenced in Agent-Ping — Kevin's one-file CLI is a concrete HITL primitive (notify/ask/confirm/choose) implementing this "stay in the loop" pattern model-agnostically. Source: User, 2026-06-17
• 2026-06-19 | Split the live-control layer into Human-in-the-Loop Control while keeping this page focused on human comprehension and teacher-mode review. Source: whole-wiki concept review, 2026-06-19