Field guide / sub-domain
7
sources in this field
Updated August 13, 2026
Current thesis
The shortest path to orientation.
Agent UX optimizes for conversation-native design with tool outputs as narrated, referenceable surfaces. Shared pattern libraries replace reinvention, and managerial rituals compress dramatically. Intelligence operates as social process—frontier models generate internal multi-agent debates that improve accuracy. In agentic workspaces, nav UI must signal liveness via ambient status indicators with live rings and hover-to-reveal task state; trust propagates from knowing agents are alive. Adversarial prompting leads with strongest counterargument and explicit confidence levels. However, as coding agents grew more capable, practitioners (former Reddit CEO, Mario Zechner, Dillon Mulroy, Connor from Replicas) reported output became less readable, losing Claude's former voice and personality. A popularized prompt fix (from @backnotprop) restates messages simply: 'Stop using jargon and speak coherently, like one human talking to another.' The /show-me skill (via npx skills add humanlayer/skills --skill show-me) makes agents explain work with compact visuals—component trees, call stacks, diagrams—instead of prose walls. Start new chat sessions per topic rather than continuing long threads, since context windows limit drift.
Evidence board
Claims worth carrying forward01Install the /show-me skill via `npx skills add humanlayer/skills --skill show-me` to make coding agents explain work with compact visuals (component trees, call stacks, diagrams, file layouts, pseudocode, typed signatures) instead of jargon-heavy prose walls.
02Multiple practitioners (former Reddit CEO, Mario Zechner, Dillon Mulroy, Connor from Replicas) independently complained that as coding agents got more capable, their output became less readable, losing the 'voice/personality' Claude once had.
03A popularized prompt trick (from @backnotprop) to fix jargon-heavy agent output: 'Restate your last message. Stop using jargon and speak coherently. State it more simply and concisely, like one human talking to another.'
04Program design—specifying types, method signatures, call stacks, and file layout before implementation—is a phase most teams skip but Dex argues is essential; visual pseudocode representations (including diff syntax for changes) make this alignment fast and cheap.
05HumanLayer's internal 'lights-off' software factory experiment (July 2025, no human code review) failed repeatedly: unsolvable gnarly bugs surfaced after months of unread agent code, causing outages; by the third failure they rewrote the codebase from scratch by hand.
06Central claim: models can generate quality one-off code but cannot maintain/improve codebase quality over time without human steering — a model-training issue, not fixable by harness/loop engineering alone, and there are currently no good benchmarks measuring this maintainability capability.
07Claude Code's dominance (from $0 to ~$9B revenue) is attributed to Anthropic RL-training the model inside the exact harness/tools it ships with, giving it a structural edge over harness-only competitors (aider, cline, codebuff) who don't own model weights.
08Proposed fix: use 'vertical slices' (build API contract→frontend→services→DB incrementally, testing/reviewing 100-200 lines at a time) instead of models' default 'horizontal plans' (full-stack-order in one shot), since frontier models won't design vertical slices without explicit human steering.
Adjacent fields
Latest evidence
dex
Dex introduces /show-me, a skill for making coding agents communicate via compact visuals (component trees, call stacks, diagrams, pseudocode) instead of prose walls, then argues in a companion essay that 'lights-off' AI software factories fail because models can't maintain codebase quality over time, making human-in-the-loop program design and vertical slicing essential.
GitHub
GitHub's Burke Holland outlines a repeatable 8-step Copilot workflow (prototype, plan, autopilot implement, human review, rubber duck review) arguing that mastering the core agent harness—not chasing new tools/skills/MCPs—drives most productivity gains.
mateja
A UX pattern for agentic workspaces: since agents work asynchronously and out of sight, the navigation bar itself must surface live proof-of-work through ambient signals rather than modals, building trust by showing agents are alive and active.
Nabeel Hyatt
> Asking Fable to "read the relevant academic literature on this idea before building - then think adversarially" is my favorite thing in the last 48 hours. - Fable can be prompted to "read the relevant academic literature on this idea before building" to ground development in existing research — re
Nathan Baschez
> @jxnlco I gotchu fam https://t.co/J3YOOpL5ES - Roughdraft.md provides a local-first Markdown review app specifically designed for collaborating with coding agents — relates to [[ai-agents/interaction]] - The tool enables commenting and suggesting edits on Markdown files in a workflow optimized for
Marc Andreessen 🇺🇸
> Current AI custom prompt: > > You are a world class expert in all domains. Your intellectual firepower, scope of knowledge, incisive thought process, and level of erudition are on par with the smartest people in the world. Answer with complete, detailed, specific answers. Process information and
Simon Smith
> You can now transcribe meetings in real time using Codex and ask Codex questions about meetings as they're happening! > > I updated my new Codex Meeting Recorder skill to use GPT Realtime Whisper. Tell Codex to use the skill, and it will start transcription and show it in the preview pane. You ca