HackerNews AI - 2026-08-17¶
1. What People Are Talking About¶
August 17's Hacker News AI feed expanded to 80 stories from 78 authors, with 1,002 total points and 892 total comments. That was a large step up from August 16's 58 stories, 467 points, and 241 comments. Attention concentrated sharply around two submissions: galnagli posted AI-Generated GitHub Copilot “Autofix” Allowed Compromise of Snowflake's Jira (281 points, 117 comments), and jacquesm posted On AI regulation and messaging (224 points, 479 comments). Together they produced about half the day's points and about two-thirds of its comments, while the top five stories drove about 74% of points and 88% of comments. The feed was still builder-heavy - 20 Show HNs and 6 Ask HNs - but the center of gravity shifted from "what can agents do?" to "who should control them, how do we verify them, and what parts of human work should they touch at all?"
1.1 Frontier AI legitimacy got argued in terms of trust, power, and openness (🡕)¶
The day's loudest conversation was about legitimacy rather than raw capability. The central dispute was not whether AI is powerful. It was whether frontier labs have earned the right to shape policy and defaults for everyone else while keeping the most powerful systems, tooling, and compute under tight control.
jacquesm posted On AI regulation and messaging (224 points, 479 comments). The linked Dario Amodei thread argues that regulation does not automatically imply capture, says Anthropic tries to support rules that slow frontier labs more than smaller competitors, and says AI structurally concentrates power because of scaling laws even when open-weights exist. The same thread says trust will come from concrete results in biology and medicine rather than from better marketing.
HN comments mostly treated that as insufficient. kilpikaarna (score 0) quoted the thread's "cure cancer" language as exactly the kind of promise people no longer trust, while mindwok (score 0) argued Anthropic's public-good rhetoric rings hollow unless it ships something materially empowering, such as open-source tooling or open-weight models. pu_pe (score 0) accepted Dario's concentration argument in principle, but still treated compute ownership as the real locus of power.
bilsbie posted Anthropic's War on open source AI (122 points, 52 comments). The linked xcancel post frames open-source AI as a freedom and private-compute issue, casting closed AI as dependence on a narrow set of institutional gatekeepers. Even when the rhetoric was extreme, it landed because it gave frustrated users a concrete counter-model to frontier-lab paternalism.
Discussion insight: HN's open-source argument is no longer just "open is nice." It is increasingly being used as the concrete proof of trust, agency, and power-sharing when frontier labs ask for deference.
Comparison to prior day: August 16 focused on multi-agent coordination and liability once autonomous systems are already deployed. August 17 moved earlier in the stack to the fight over who gets to hold power in the first place and what kind of AI infrastructure technical users will actually accept.
1.2 Verification and execution-path safety became the practical counterweight to cheap AI-generated code (🡕)¶
The second dominant thread made AI-assisted software risk painfully concrete. The point was not that insecure code exists. It was that AI lowers the cost of introducing changes much faster than it lowers the cost of safely reviewing or constraining them.
galnagli posted AI-Generated GitHub Copilot “Autofix” Allowed Compromise of Snowflake's Jira (281 points, 117 comments). The linked Wiz write-up says Wiz Red Agent found a GitHub Actions script-injection vulnerability after a PR co-authored by "Copilot Autofix powered by AI" replaced a safe env plus jq pattern with direct issue-title interpolation. Wiz then exploited the workflow by opening a crafted GitHub issue, exfiltrated a Jira token from the runner, and triggered same-day remediation and credential rotation.
The most useful HN responses immediately converted the incident into operating guidance. inahga (score 0) said GitHub Actions should be statically analyzed with tools such as zizmor, and CodeWithLeo (score 0) said the real lesson is that verification, not code generation, is now the bottleneck. mjr00 (score 0) pushed the same point from an internal-process angle: AI makes it cheap to spin up low-value changes that are still expensive for humans to understand, review, and maintain.
Lower in the review set, builders were already turning the same lesson into product surfaces. alanfuNZ posted Show HN: Doberman: The AI watchdog that stops Claude from deleting your database (4 points, 1 comment). The Doberman README says it sits directly on the tool-execution path as an MCP proxy or host hook and forces each action through a PASS, AUTH, or BLOCK verdict before the call runs, explicitly arguing that anything off the execution path is advisory rather than protective.
Discussion insight: The community response was not anti-AI fatalism. It was to add narrower, more mechanical gates: static analysis, fail-closed approval layers, and human review that assumes generated code is cheap but not trustworthy.
Comparison to prior day: August 16 argued for smaller diffs, proof receipts, and command guards. August 17 supplied a concrete public exploitation story showing exactly why those measures are becoming table stakes.
1.3 Builders kept packaging routing, protocol, and operator surfaces instead of one more model (🡕)¶
Builder energy kept moving up a layer from single models to coordination surfaces. The most interesting launches were not new frontier models. They were routers, standards, galleries, and editing layers that make complex agent systems easier to swap, supervise, or correct.
abdik posted Launch HN: Speko (YC S26) – OpenRouter for Voice AI (80 points, 50 comments). The selftext describes production voice agents as three-layer ensembles of STT, LLM, and TTS models that teams rarely re-evaluate because switching vendors looks like another R&D project. Speko's pitch is a public benchmark plus routing layer, while the open-source Speko Gateway keeps BYOK credentials local and can optionally use managed routing on top.
songrenchu posted Show HN: HarnessRouter: Unified interface for agent harnesses (6 points, 9 comments). The selftext and the Unified Harness Protocol both argue that the missing abstraction is not another agent loop but a standard task API for driving whole harnesses such as Codex, Claude Code, and Hermes. The HarnessRouter README pushes the same idea operationally: one local container, your own API keys, your own data, and files or sessions back instead of raw token streams.
engomez posted Show HN: Visimer – open-source visual editor for Mermaid diagrams (13 points, 0 comments), and both the selftext and README emphasize point-editing AI-generated diagrams without rewriting the whole artifact. nastynate posted Show HN: A community library for Claude Code status lines (10 points, 1 comment), and statuslin.es says every submitted status-line script is sandbox-rendered once, human-reviewed, and then served as a pinned preview instead of executing strangers' shell code on page view. miketromba posted Show HN: Cronloop, run Claude Code or Codex on a schedule (2 points, 1 comment), positioning recurring agent runs and persistent memory as another missing operator surface.
Discussion insight: Across very different products, the repeated move was to externalize control - benchmark boards, standardized task APIs, point-edit artifacts, safe preview galleries, and scheduled loops - rather than ask users to trust a generic chat window.
Comparison to prior day: August 16 emphasized local-first memory, MCP tools, and native workbenches. August 17 widened that pattern into routing and protocol layers that try to make entire agent stacks swappable.
1.4 Creative AI tools kept colliding with authorship and meaning (🡕)¶
Some of the day's most animated discussion was not about enterprise agents at all. It was about whether AI tools can participate in creative work without hollowing out authorship, motivation, or the economic basis for doing the work.
refsab posted Show HN: 1667, a terminal UI for writing fiction with language models (33 points, 90 comments). Its selftext and site are careful about the boundary: the model writes only when asked, each take stays in a local story tree, there is no telemetry, and "the judgement stays yours." But the HN replies were often hostile to the premise itself. DC-3 (score 0) argued that having the model write the prose voids the aspiration of writing, while dbspin (score 0) made the same point more bluntly. By contrast, thegagne (score 0) wanted a nearby product: story-bible management, prompts, critique, and editorial feedback without auto-authored text.
bg117 posted Ask HN: Is this the worst time to be human? (16 points, 44 comments), asking whether AI will strip developers and artists of both joy and economic leverage. The stronger replies pushed back on the historical framing, but dgllghr (score 0) agreed with the underlying signal: AI makes the present feel unusually uncertain because multiple cultural and professional anchors are shifting at once.
solomonyardley posted Show HN: Tesana – An AI game engine that builds quality games end-to-end (7 points, 3 comments), describing a prompt-to-game platform with 250,000 builders and web publishing. That mattered less for the comments than for the contrast: one creative product carefully subordinates the model to the human, while another sells end-to-end generation as the main convenience.
Discussion insight: HN was not uniformly anti-creative-AI. The boundary it kept drawing was between tools that preserve human judgment and tools that appear to replace the act people actually value.
Comparison to prior day: August 16 was overwhelmingly about agent infrastructure and verification. August 17 pulled the consequences back into creative and emotional territory.
2. What Frustrates People¶
Frontier AI companies keep asking for trust they have not yet earned¶
jacquesm posted On AI regulation and messaging (224 points, 479 comments), and the linked Dario Amodei thread argues trust should come from actual public benefit rather than better marketing. HN's answer was that messaging is still too top-down. kilpikaarna (score 0) said the promises ignore current concerns such as employment, energy, and concentration, while mindwok (score 0) asked for open models or open tooling instead of more rhetoric. bilsbie posted Anthropic's War on open source AI (122 points, 52 comments), and its linked thread turns the same frustration into a blunt demand for open source and private compute. The frustration is that legitimacy claims still arrive as assurances, while the audience wants inspectable control. Severity: High. Worth building for: yes, competitively.
Human review and CI hardening lag behind AI-generated change¶
galnagli posted AI-Generated GitHub Copilot “Autofix” Allowed Compromise of Snowflake's Jira (281 points, 117 comments), and the linked Wiz report shows how a workflow change co-authored by Copilot Autofix removed a safe pattern and opened a live GitHub Actions injection path. inahga (score 0) immediately suggested zizmor in CI, while CodeWithLeo (score 0) said the real bottleneck has shifted from generation to verification. alanfuNZ posted Show HN: Doberman: The AI watchdog that stops Claude from deleting your database (4 points, 1 comment), and its README exists for the same reason: prompt-level guardrails are not enough if dangerous calls still reach the execution path. The frustration is not just insecure code. It is the widening gap between how cheap AI makes change and how expensive humans still find safe review. Severity: High. Worth building for: yes, directly.
Comparing or switching multi-model stacks is still too much manual integration work¶
abdik posted Launch HN: Speko (YC S26) – OpenRouter for Voice AI (80 points, 50 comments) because teams usually benchmark voice stacks once and then stop revisiting them, even as better STT, LLM, and TTS options appear. The comments from webo (score 0) and spmartin823 (score 0) asked for trustworthy measurement and a more unified "conversation in a box" surface, while MikhailTal (score 0) immediately compared the product to LiveKit and Vapi. songrenchu posted Show HN: HarnessRouter: Unified interface for agent harnesses (6 points, 9 comments) because harnesses themselves now have incompatible request and session models. The frustration is that fragmentation keeps forcing teams to re-solve routing, switching, and governance as one-off engineering projects. Severity: High. Worth building for: yes, directly.
People want creative help without giving up authorship or joy¶
refsab posted Show HN: 1667, a terminal UI for writing fiction with language models (33 points, 90 comments), and the site goes out of its way to say the model only writes when asked and that the writer's judgment stays in charge. That still was not enough for many readers. DC-3 (score 0) and dbspin (score 0) argued that model-written prose undermines the point of writing, while thegagne (score 0) described a nearby unmet need: planning, critique, story-bible management, and feedback without auto-authored text. bg117 posted Ask HN: Is this the worst time to be human? (16 points, 44 comments), turning the same anxiety toward coding and visual art. The frustration is both practical and emotional: people want assistance, but not at the cost of meaning, craft, or ownership. Severity: Medium-High. Worth building for: yes, competitively.
Small local conventions still shape whether agent tools feel respectful¶
joooscha posted Pi coding agent: config folder is out of place on Linux (45 points, 19 comments), and the HN thread turned a seemingly small XDG-path complaint into a product-choice signal. Systemerror7A69 (score 0) said the issue was frustrating enough to make them explore other coding agents, while joshka (score 0) framed OS-specific config placement as a reasonable expectation rather than bikeshedding. On the positive side, nastynate posted Show HN: A community library for Claude Code status lines (10 points, 1 comment), and statuslin.es exists because people care enough about daily ergonomics to browse, preview, and safely copy terminal UI details. The frustration is that agent tools often ask for deep trust while still mishandling basic local workflow expectations. Severity: Medium. Worth building for: yes, competitively.
3. What People Wish Existed¶
Openness and local control that function as proof, not branding¶
The clearest trust request in the dataset was not "better messaging." It was concrete user control. jacquesm posted On AI regulation and messaging (224 points, 479 comments), and the linked Dario Amodei thread explicitly argues that AI structurally concentrates power and that fair institutions can constrain that. HN's response from mindwok (score 0) was to ask for open models or open tooling, while bilsbie posted Anthropic's War on open source AI (122 points, 52 comments), arguing for open source and private compute as the alternative. The unmet need is for products whose openness and local control actually demonstrate alignment with users instead of merely claiming it. This is a practical need, and the opportunity is direct-to-competitive.
Verification layers that fail closed before bad agent actions land¶
galnagli posted AI-Generated GitHub Copilot “Autofix” Allowed Compromise of Snowflake's Jira (281 points, 117 comments), and the linked Wiz write-up shows why AI-written changes need stronger mechanical gates than ordinary PR review. inahga (score 0) called for workflow static analysis, and alanfuNZ posted Show HN: Doberman: The AI watchdog that stops Claude from deleting your database (4 points, 1 comment), whose README makes execution-path enforcement the product itself. The need is for systems that verify, lint, or block before a merge or tool call, not after the damage is already live. This is a practical and immediate need, and the opportunity is direct.
Vendor-neutral routing for voice stacks and agent harnesses¶
abdik posted Launch HN: Speko (YC S26) – OpenRouter for Voice AI (80 points, 50 comments) because STT, LLM, and TTS choices keep changing faster than most teams can re-benchmark or re-integrate them. songrenchu posted Show HN: HarnessRouter: Unified interface for agent harnesses (6 points, 9 comments) because Codex, Claude Code, Hermes, and similar harnesses still expose incompatible control surfaces. miketromba posted Show HN: Cronloop, run Claude Code or Codex on a schedule (2 points, 1 comment), extending the same request into recurring automation. The unmet need is a durable control plane above rapidly shifting providers and runtimes. This is a practical need, and the opportunity is direct.
Creative copilots that organize, critique, and preserve human judgment¶
refsab posted Show HN: 1667, a terminal UI for writing fiction with language models (33 points, 90 comments), and the site deliberately keeps the human in charge of selection and export. But thegagne (score 0) immediately described a different desired product: planning help, feedback, prompts, and continuity management without model-authored prose. engomez posted Show HN: Visimer – open-source visual editor for Mermaid diagrams (13 points, 0 comments), and the README embodies the same instinct for diagrams: let AI help generate the artifact, then give the human precise point-edit control. The need is not just "AI for creativity." It is AI that helps without seizing the part of the process people identify with. This is a practical and emotional need, and the opportunity is competitive.
Agent workbenches that feel native, inspectable, and schedulable¶
joooscha posted Pi coding agent: config folder is out of place on Linux (45 points, 19 comments) because even a path convention can signal whether a tool respects a user's environment. nastynate posted Show HN: A community library for Claude Code status lines (10 points, 1 comment), and statuslin.es turns small daily UI improvements into a safely previewed gallery. miketromba posted Show HN: Cronloop, run Claude Code or Codex on a schedule (2 points, 1 comment), showing the same demand for operational surfaces that fit real routines. The need is for agent tools that feel like first-class local software rather than remote demos with rough edges. This is a practical need, and the opportunity is competitive.
4. Tools and Methods in Use¶
| Tool | Category | Sentiment | Strengths | Limitations |
|---|---|---|---|---|
| GitHub Copilot Autofix | Coding assistant / security review | (-) | Makes low-cost workflow edits and participates in review flows | Missed a critical template-injection flaw and replaced a safer env plus jq pattern |
| Wiz Red Agent | AI security agent | (+) | Rapidly finds and validates exploitable CI/CD weaknesses | Helps defenders discover issues only after exposure exists; compresses patch windows |
zizmor |
CI static analysis | (+) | Specifically catches GitHub Actions template-expansion problems | Only helps if teams wire it into CI and treat findings as blocking |
| Speko / Speko Gateway | Voice AI routing | (+/-) | Public benchmarks, route selection, BYOK local gateway, easier vendor switching | Benchmark trust, turn-taking, and differentiation versus LiveKit or Vapi remain open questions |
| HarnessRouter / UHP | Agent orchestration | (+) | One task-oriented surface for multiple harnesses, local runtime, open protocol, starter kits | Draft standard, fragmented ecosystem, and still-early adoption |
| Visimer | Diagram editing | (+) | Lets humans refine AI-generated Mermaid artifacts with minimal, lossless edits | Narrow surface area; most valuable only in Mermaid-heavy workflows |
| 1667 | Creative writing tool | (+/-) | Local-first branching story tree, provider choice, no telemetry, human-controlled export | Strong authorship backlash and a deliberately niche audience |
| statuslin.es | Claude Code UX | (+) | Safe preview gallery, sandboxed rendering, and human review before publish | Narrow ergonomics win rather than a broad platform layer |
| Doberman | Agent guardrails | (+) | Execution-path PASS/AUTH/BLOCK model, fail-closed behavior, works with Claude Code, Codex, and MCP | Alpha status and added approval friction on sensitive operations |
| Cronloop | Scheduled agent automation | (+) | Repeating runs, live monitoring, and memory between runs | Little evidence yet on governance and reliability at larger scale |
Overall sentiment was strongest for tools that narrowed or explained the trust boundary: static analyzers, execution-path guardrails, public benchmark boards, local or self-hosted runtimes, and safe preview surfaces. The mixed or negative cases were the ones that asked users to accept more automation without giving them more evidence or control.
The common workarounds were strikingly mechanical. Add CI linting. Keep credentials local. Standardize harness interfaces. Preserve exact text diffs. Preview arbitrary scripts in a sandbox. Schedule agents as explicit jobs instead of leaving them as ad hoc chat sessions. The market is repeatedly choosing products that make agent behavior easier to inspect, replay, or swap out.
Competitive dynamics are splitting into three layers. One layer is proof and safety for AI-written code. Another is routing and orchestration above fragmented model or harness ecosystems. The third is edit-first or operator-first surfaces that decide how much judgment stays with the human.
5. What People Are Building¶
| Project | Who built it | What it does | Problem it solves | Stack | Stage | Links |
|---|---|---|---|---|---|---|
| Speko | abdik | Routes across benchmarked STT, LLM, and TTS combinations for voice agents | Teams rarely re-evaluate voice stacks because switching vendors is operationally expensive | Hosted router, public benchmark boards, Go gateway sidecar, BYOK local socket, LiveKit integration | Beta | post, site, gateway |
| 1667 | refsab | Terminal UI for long-form fiction with branching takes and local project files | Creative writers want model assistance without giving up local control or structure | Terminal app, local file tree, OpenAI- and Anthropic-compatible endpoints, Ollama, LM Studio, llama.cpp | Beta | post, site |
| Visimer | engomez | Visual editor for Mermaid diagrams that keeps source and canvas in sync | AI-generated or hand-written Mermaid diagrams are tedious to fine-tune in raw syntax | TypeScript packages, Mermaid.js, SVG-to-CST mapping, Monaco, CodeMirror | Beta | post, repo, playground |
| HarnessRouter | songrenchu | Runs agent harnesses such as Codex, Claude Code, and Hermes behind one API | Products do not want to rebuild orchestration separately for every harness | Docker, UHP, local self-hosting, task/event/session API, starter kits | Beta | post, repo, UHP |
| statuslin.es | nastynate | Community gallery of Claude Code status lines with rendered previews | People want to customize daily agent UX without running arbitrary shell code from strangers | Bun, React SSR, Drizzle, Postgres, E2B sandboxing, GitHub auth | Shipped | post, site, repo |
| Doberman | alanfuNZ | Runtime guardrail that decides PASS, AUTH, or BLOCK before an agent tool call executes | Prompt-level safety is not enough once agents can run destructive or exfiltrating actions | Python, MCP proxy, host hooks, runtime risk engine, approval flow | Alpha | post, site, repo |
| Cronloop | miketromba | Runs Claude Code or Codex on a recurring schedule with memory between runs | Recurring agent work is awkward to manage as one-off terminal sessions | Scheduled loops, Markdown job definitions, run history, persistent memory | Beta | post, site |
| Tesana | solomonyardley | Prompt-to-game platform that plans, builds, and publishes browser games | Non-programmers want an end-to-end path from idea to playable game | AI planning and generation loop, graphics and logic generation, web publishing | Shipped | post, site |
Speko and HarnessRouter are the clearest "route around fragmentation" bets in the dataset. One unifies voice-model stacks, the other unifies full agent harnesses, and both place their durable product value above the rapidly changing providers underneath.
Visimer, statuslin.es, and Doberman are solving a different trust problem. Visimer turns generated artifacts back into precise human-editable objects, statuslin.es makes daily customization safer to inspect and copy, and Doberman refuses to trust prompt-layer guardrails unless the runtime can actually stop a bad call. Different surfaces, same instinct: users trust agents more when behavior is visible and bounded.
1667, Cronloop, and Tesana widen the picture. 1667 sells creative assistance with local control and explicit human judgment, Cronloop treats repeated agent work as a schedulable operational surface, and Tesana pushes the opposite direction toward faster end-to-end generation. That split helps explain why August 17 mixed strong builder energy with unusually direct anxiety about authorship and control.
6. New and Notable¶
Wiz turned AI-assisted CI/CD failure into a concrete exploitation story¶
galnagli posted AI-Generated GitHub Copilot “Autofix” Allowed Compromise of Snowflake's Jira (281 points, 117 comments). The linked Wiz report is notable because it compresses the whole AI software-risk argument into one public example: an AI-co-authored workflow change removed a safer pattern, a security agent exploited the result within days, and the fix required old-fashioned remediation, rotation, and auditing.
Openness became the day's main battlefield for AI legitimacy¶
jacquesm posted On AI regulation and messaging (224 points, 479 comments), and bilsbie posted Anthropic's War on open source AI (122 points, 52 comments). Together they were notable because they showed the trust debate hardening into a control debate: not merely whether AI should be safe, but whether it should remain closed and centrally mediated by a few labs.
Voice AI routing looked increasingly like a product category of its own¶
abdik posted Launch HN: Speko (YC S26) – OpenRouter for Voice AI (80 points, 50 comments). The product is notable because it treats benchmarked model selection, switching, and local credential control as the thing being sold, which suggests the enduring value may sit above any specific STT, LLM, or TTS provider.
A local-first fiction tool triggered one of the sharpest authorship debates in the dataset¶
refsab posted Show HN: 1667, a terminal UI for writing fiction with language models (33 points, 90 comments). It was notable because even a deliberately constrained, no-telemetry, human-in-control writing tool still triggered strong resistance from people who saw model-written prose as incompatible with the act of writing itself.
7. Where the Opportunities Are¶
[+++] Verification and execution-path safety for AI-generated code and CI - The Wiz incident, calls for zizmor, and Doberman's runtime guardrails all point to the same gap. AI is already cheap enough to create risky workflow changes faster than humans can safely review them, so products that lint, prove, or block before execution have immediate, concrete value.
[+++] Open, user-controlled AI infrastructure that turns trust into something inspectable - The Dario thread backlash and the open-source AI thread both show that "trust us" messaging has weak traction with technical users. Products that keep keys local, run on the user's machine, or make openness operational rather than rhetorical have a strong opening.
[++] Routing and orchestration layers for fragmented voice and harness ecosystems - Speko, HarnessRouter, and Cronloop all attack the same structural problem from different angles: too many model combinations, too many runtime interfaces, and too much repeated glue code. The opportunity is moderate-to-strong because the pain is practical, but several credible approaches are already forming.
[++] Edit-first creative tools that preserve authorship while still using models - 1667 and Visimer show a path where AI generates drafts or structure but the human still owns judgment and fine edits. The signal is strong enough to matter, but the cultural resistance around replacement means product design has to be careful.
[+] Daily-use agent ergonomics and operator surfaces - The Pi config-folder thread, statuslin.es, and scheduled-loop tooling suggest a secondary opportunity in the mundane layer: local conventions, visibility, repeatability, and workflow polish. This is an emerging signal, but users are clearly willing to switch tools over it.
8. Takeaways¶
- AI legitimacy is increasingly being judged through power and openness, not just benchmark talk. The largest discussion thread on the day turned quickly from regulation details to whether frontier labs have earned trust without sharing enough control. (source)
- AI made software change cheaper faster than it made software review safer. The Snowflake workflow incident mattered because it showed an AI-co-authored change opening a real exploit path that another AI system found within days. (source)
- Routing and orchestration are becoming durable product layers above fast-changing model ecosystems. Speko and HarnessRouter both sell selection, switching, and control surfaces rather than a specific base model. (source)
- Users keep rewarding tools that narrow the trust boundary with explicit artifacts, previews, or rules. Visimer, statuslin.es, and Doberman all succeed by making AI behavior easier to inspect, edit, or block. (source)
- Creative AI remains acceptable mainly when people feel the judgment still belongs to them. The 1667 discussion made clear that many users will accept organization, prompting, or critique long before they accept model-authored creative output as the main act. (source)