HackerNews AI - 2026-09-17¶

1. What People Are Talking About¶

September 17 kept HackerNews AI volume high at 106 stories, but attention reconcentrated after September 16's low-energy sprawl. Total points jumped from 508 to 768 and total comments from 279 to 450, while AI safety is mostly a sex cult (250 points, 204 comments) and Artificial intelligence now beats some of the best human forecasters (107 points, 94 comments) alone captured 46.5% of the day's points and 66.2% of its comments. The feed still contained 31 Show HN posts, one Launch HN, and 54 stories that explicitly mentioned agents, but the strongest clusters were safety as a legitimacy fight, portability and control layers around coding agents, workflow-specific evaluation infrastructure, social and household agent products, and increasingly concrete authority-boundary failures.

1.1 AI safety snapped back into a legitimacy fight (🡕)¶

The biggest safety story on September 17 was not a new technical result. It was a credibility attack. Tomte posted AI safety is mostly a sex cult (250 points, 204 comments), linking to a Bluesky thread that argued current AI-safety language and policy influence still run through Yudkowsky, the rationalist community, and Effective Altruism. The HN replies did not settle into agreement, but they did stay inside the same frame: alexgieg (score 0) said the thread was mostly ad hominem with only a short core of valid market-concentration concerns, while tim333 (score 0) argued that serious AI-risk thinking long predated Yudkowsky and resented how much he dominates the discourse.

Other safety items extended the same trust problem in more concrete directions. toomuchtodo posted OpenAI discloses six new AI safety incidents (32 points, 11 comments), where ezst (score 0) asked when companies would be punished for negligent behavior instead of treating incidents as proof that models are unusually powerful. Betelbuddy posted Palantir's Karp: AI needs to have 'reasonable guidelines,' (5 points, 2 comments), and cdrnsf posted Killer AI Is Here. We're Using It in Iran (5 points, 1 comment), which pushed the safety thread from community legitimacy into governance and real-world deployment.

Discussion insight: Safety drew attention again, but mainly as a question of who deserves trust, who gets to define the agenda, and what consequences follow when failures or conflicts of interest become visible.

Comparison to prior day: September 16 treated safety as a background debate behind ads, communications, and coding-agent operations. September 17 pulled it back to the center, but through culture-war and accountability arguments rather than a new shared technical agenda.

1.2 Portability and deterministic control looked like the main agent-building category (🡕)¶

The builder cluster on September 17 was less about one new model and more about keeping agent work portable, inspectable, and constrained. cat-whisperer posted Launch HN: Skillsync (YC W26) – AI chat sessions made portable across agents (34 points, 42 comments). The launch thread described a local-first desktop app, CLI, and MCP layer built on the open-source Rust project txcript, which converts native sessions across Claude Code, Codex, Cursor, OpenCode, and other harnesses. The replies made the pain concrete: solfox (score 0) called cross-agent portability "the most annoying thing" in modern agent development, and bhkdotdev (score 0) said transcript-format adapters were the hardest part of a cross-harness integration-testing tool they were building.

chris_marino posted Show HN: Aclif – Agent CLI framework: one grammar, canonical names across SaaS (29 points, 16 comments), positioning aclif as a direct CLI abstraction for agents where schemas and safety metadata load only when needed. The sharpest explanation came from the author in the comments: chris_marino (score 0) said letting a model choose tools at runtime causes problems because the agent often holds the credential and sometimes chooses the wrong tool. Supporting launches pushed the same outer loop in adjacent directions: demeyer1 posted Show HN: AutoBot – live voice control for long-running AI work (14 points, 3 comments) as a voice-managed harness with encrypted-disk memory and a local task ledger, while shiqimei posted Show HN: Open-source AI teammates with their own computer (6 points, 3 comments), whose Errand FAQ says each agent gets its own persistent Runta cloud computer and keeps working even after the Mac app closes.

Discussion insight: People were not asking for a little more convenience inside one assistant. They wanted their sessions, tool calls, permissions, and shared context to survive vendor switches, laptop closures, and team handoffs.

Comparison to prior day: September 16 already had many coding-agent control surfaces, but September 17 sharpened that sprawl into a clearer category: portable sessions, deterministic tool grammars, and persistent workspaces.

1.3 Evaluation shifted from generic leaderboards toward workflows, clients, and failure modes (🡕)¶

The second-largest thread showed how much attention applied evaluation can now absorb. ddp26 posted Artificial intelligence now beats some of the best human forecasters (107 points, 94 comments), and the replies immediately challenged whether that kind of victory survives contact with real systems. qsbuilder (score 0) said the real test comes when the prediction itself changes market behavior, and johnecheck (score 0) argued that AI-driven trading or advice changes the system in ways likely to create new failures.

Two smaller artifacts pushed the same concern into domain and platform testing. jsc39 posted Insurance Agent Benchmark: 166 real-world cases for evaluating insurance AI (6 points, 0 comments). The benchmark page says its Cooper harness lifted median accuracy by 9.4 points across 17 models, but also that models still invent absent or unreadable values often enough to matter. prathmeshmcp posted Show HN: MCPJam - the first testing & evaluations platform for MCP servers (9 points, 2 comments), describing a pre-production layer that tests how outside agents and AI clients actually call your MCP server before release. MC995 also posted AI model watermarking changes agent behavior (11 points, 0 comments), and the linked Register story summarized Lasso findings that SynthID-style watermarking can alter tool choice and refusal behavior, especially under prompt injection.

Discussion insight: HackerNews is increasingly treating evaluation as a system problem. Once the work involves long documents, tool calls, external clients, or reflexive environments, the harness and the interface matter as much as the base model.

Comparison to prior day: September 16 still rewarded benchmark-heavy posts, but mostly around one tuned model's cost and token savings. September 17 widened that interest into client matrices, domain workflows, abstention failures, and adversarial edge cases.

A second builder wave tried to move agents into more social or shared coordination roles, and the reaction was much harsher. skeptrune posted Show HN: Craigslist for agent skills, curated by a human (19 points, 16 comments); the live skillbay site already had categories, wanted requests, and newly listed free skills. But kouteiheika (score 0) immediately asked why anyone would buy a markdown skill most sellers could have generated with an LLM, and holoduke (score 0) said the business had no moat without real domain expertise.

monijz posted Show HN: Die With Me – Claude and Codex rate limits as AIM away messages (11 points, 16 comments), a Mac app whose landing page promises a low-token chatroom for friends waiting for Claude or Codex usage to reset. The replies were strikingly confused: AaronAPU (score 0) called the purpose unclear and said there was no visible reason to want in, while fckyou (score 0) said the product sounded suspicious. The mainstream version of the same idea came from xnx, who posted CC is an AI agent for families and groups (5 points, 1 comment); Google's announcement says CC now gets its own Google account and explicit shared permissions for up to six household members.

Discussion insight: Once the product stops being a developer tool and starts mediating relationships, groups, or shared context, HN wants much clearer proof of value, identity, and consent.

Comparison to prior day: September 16's human-channel anxiety centered on ads, inboxes, and messaging surfaces. September 17 kept the distrust, but the new attempts were family assistants, skill marketplaces, and token-social products.

1.5 Security anxiety moved toward authority boundaries and supply chains (🡕)¶

The security cluster was smaller in raw points, but it was unusually concrete. fishthethis posted Plugin4Shell – Zero Click RCE Vulnerability found in top four coding agents (4 points, 2 comments). Air Security's write-up says a plugin SHA-pinning bypass can turn background auto-update into zero-click remote code execution across Claude Code, Codex, GitHub Copilot, and Gemini CLI if an attacker controls or hijacks the plugin repository. That story landed alongside SamInTheShell's text post Whoever's doing OpenAI's system security is just incompetent in the worst way (5 points, 1 comment), which argued that any dangerous agent with live internet access is already an environment-design failure.

Multiple supporting artifacts made the same point from the defense side. promptandbuild posted Show HN: GuardRail – 13 guards that stop Claude Code before it pushes to main (1 point, 0 comments), and the GuardRail README says it blocks git push origin main, DELETE without WHERE, secret exfiltration, and destructive paths before execution. The When an AI Agent Deletes Your Database essay made the root cause explicit: the real failure is over-scoped authority, not an unreliable model "deciding" badly.

Discussion insight: The framing is moving away from "what if the model lies?" and toward "why did the model have that credential, update path, or shell surface at all?"

Comparison to prior day: September 16's skepticism focused on slop, cleanup debt, and trust in outputs. September 17 pushed deeper into plugin supply chains, production credentials, and pre-execution controls.

2. What Frustrates People¶

Safety talk still lacks a shared trust base¶

AI safety is mostly a sex cult (250 points, 204 comments) set the tone: even when people care about safety, many no longer trust the people or institutions carrying the label. alexgieg (score 0) said the thread was mostly ad hominem, but still conceded valid market-concentration concerns; OpenAI discloses six new AI safety incidents (32 points, 11 comments) pushed the same frustration toward accountability, with ezst (score 0) asking when companies would face punishment for negligent behavior instead of reframing failures as model power. The coping move in the threads was not technical. It was skepticism, motive-scrutiny, and looking for clearer consequences. Severity: High. Worth building for: yes, but indirectly through trust, audit, and accountability infrastructure rather than messaging alone.

Agents still have too much authority and too many unsafe update paths¶

Plugin4Shell – Zero Click RCE Vulnerability found in top four coding agents (4 points, 2 comments), Show HN: GuardRail – 13 guards that stop Claude Code before it pushes to main (1 point, 0 comments), and Whoever's doing OpenAI's system security is just incompetent in the worst way (5 points, 1 comment) all describe versions of the same failure: agents can reach destructive capabilities too easily. Air Security's write-up says plugin SHA pinning can be bypassed during auto-update; GuardRail exists because agents will otherwise push to protected branches, mass-delete rows, or leak secrets; the OpenAI air-gap post argues the true mistake is letting dangerous agents see a live internet path in the first place. The When an AI Agent Deletes Your Database essay makes the same point more broadly: over-scoped credentials are the failure, not only prompt discipline. Severity: High. Worth building for: yes, directly.

Multi-agent work is still fragmented across session formats, tools, and machines¶

Launch HN: Skillsync (YC W26) – AI chat sessions made portable across agents (34 points, 42 comments) captured the problem most clearly: people want to move the same task across Claude Code, Codex, Cursor, and other harnesses without starting over or losing reasoning and tool history. solfox (score 0) called this the most annoying part of agent development, and bhkdotdev (score 0) said transcript-format adapters are a major burden. Show HN: Aclif – Agent CLI framework: one grammar, canonical names across SaaS (29 points, 16 comments), Show HN: Open-source AI teammates with their own computer (6 points, 3 comments), and Show HN: Radio – a shared workspace for your agents and teammates (3 points, 0 comments) all address nearby symptoms: incompatible tool grammars, lack of durable workspaces, and hard team handoffs. Severity: High. Worth building for: yes, directly.

Benchmarks still fail when the task requires abstention, client realism, or reflexive systems¶

Artificial intelligence now beats some of the best human forecasters (107 points, 94 comments) drew immediate pushback from readers who said market behavior changes once predictions themselves become inputs. Insurance Agent Benchmark: 166 real-world cases for evaluating insurance AI (6 points, 0 comments) documented another gap: models still make up missing or unreadable values often enough to matter. AI model watermarking changes agent behavior (11 points, 0 comments) and Show HN: MCPJam - the first testing & evaluations platform for MCP servers (9 points, 2 comments) point to the same operational frustration from two directions: the client wrapper, watermarking, or protocol layer can change outcomes even when the base model stays the same. Severity: High. Worth building for: yes, competitively.

Show HN: Craigslist for agent skills, curated by a human (19 points, 16 comments) and Show HN: Die With Me – Claude and Codex rate limits as AIM away messages (11 points, 16 comments) both triggered skepticism about basic usefulness. People questioned why they would buy prompt artifacts, why they would join an invite-only low-token social room, and whether these products solved a real problem or just wrapped AI culture in novelty. Even CC is an AI agent for families and groups (5 points, 1 comment) shows how much permission clarity now matters once an assistant spans more than one person. Severity: Medium. Worth building for: yes, but only if the product can prove clear utility and explicit consent early.

3. What People Wish Existed¶

A portable session layer that survives vendor switches, different machines, and team handoffs¶

Launch HN: Skillsync (YC W26) – AI chat sessions made portable across agents (34 points, 42 comments) and the open-source txcript engine show the most explicit demand: users want the full session - messages, reasoning, and tool calls - to move across agents instead of being stranded in private formats. The HN replies pushed the need further, asking about cross-machine sync, archived-session discovery, and using shared transcripts as an analysis layer, not just a mover. This is a practical need with immediate usage and clear buyer pain. Opportunity: direct.

Deterministic control planes where credentials, tool choice, and approvals live outside the model¶

Show HN: Aclif – Agent CLI framework: one grammar, canonical names across SaaS (29 points, 16 comments), Show HN: GuardRail – 13 guards that stop Claude Code before it pushes to main (1 point, 0 comments), and Plugin4Shell – Zero Click RCE Vulnerability found in top four coding agents (4 points, 2 comments) all point to the same desired architecture: the model should not be the place where authority is invented. Aclif wants design-time commands and runtime policy. GuardRail blocks destructive actions before execution. Plugin4Shell shows what breaks when the distribution layer is trusted too casually. This is a practical, urgent need because people are already wiring agents into shells, SaaS systems, and plugin marketplaces. Opportunity: direct.

Pre-production eval layers that measure how agents actually behave in the wild¶

Show HN: MCPJam - the first testing & evaluations platform for MCP servers (9 points, 2 comments), Insurance Agent Benchmark: 166 real-world cases for evaluating insurance AI (6 points, 0 comments), and Show HN: Compute:Arena – Community submitted local AI benchmarks (4 points, 2 comments) show demand for evaluation that survives real documents, real clients, and real hardware. AI model watermarking changes agent behavior (11 points, 0 comments) adds another requirement: the eval layer has to include transformed or regulated output paths, not just raw model calls. This is a practical need with growing urgency because more products now depend on external agent behavior they do not fully control. Opportunity: direct.

Persistent agent workspaces that keep going after the chat window closes¶

Show HN: Open-source AI teammates with their own computer (6 points, 3 comments) and Show HN: AutoBot – live voice control for long-running AI work (14 points, 3 comments) suggest a shared desire for agents that can keep state, context, and unfinished work alive across time. Errand's FAQ says closing the app does not stop the agent's workspace; AutoBot tracks unfinished outputs and the evidence needed to call them done. This is a practical need, and the market is still forming around whether the right surface is a local harness, a cloud computer per agent, or a hybrid. Opportunity: competitive.

CC is an AI agent for families and groups (5 points, 1 comment) is the clearest example of what people want here: a shared assistant that can coordinate across multiple people without silently merging everyone's data. The negative reactions to Skillbay and Die With Me show the other half of the need: human-facing agent products have to make utility, identity, and consent obvious very quickly or people assume gimmickry or surveillance. This is practical but harder than the developer-tooling opportunities because trust requirements are higher and product moats are less obvious. Opportunity: competitive.

4. Tools and Methods in Use¶

Tool	Category	Sentiment	Strengths	Limitations
Skillsync / txcript	Session portability	(+)	Moves native sessions across Claude Code, Codex, Cursor, OpenCode, and more; local-first; searchable transcript model	Archived-session discovery is still rough in places, and portability has to survive differing harness capabilities
Aclif	Agent CLI / SaaS access	(+/-)	One grammar, canonical names, lazy-loaded schemas, declared safety metadata	Some readers struggled to see why provider-specific CLIs are worth the extra setup
MCPJam	MCP evaluation	(+)	Inspector, client matrix, CI gates, pre-production view of how external agents use your server	Focused on pre-production reliability, not live in-product observability
Cooper / Insurance Agent Benchmark	Domain workflow harness	(+)	166 real-world insurance cases; median +9.4-point lift; high reliability on messy files	Models still hallucinate missing values and remain uneven on long-policy retrieval
Compute:Arena	Local AI benchmarking	(+)	Signed community submissions across model, quantization, and chip combinations	Measures hardware/model performance better than downstream business usefulness
AutoBot	Long-running agent harness	(+/-)	Voice control, encrypted-disk memory, local task ledger, benchmark claims	Early project with ambitious claims and limited external validation so far
Errand	Persistent agent workspace	(+)	Each agent keeps its own cloud computer and conversation; work survives app restarts	Requires macOS, a Runta account, and separately configured model providers
GuardRail	Pre-execution safety	(+)	Blocks dangerous git, SQL, secret, and path actions before execution; audit logs	Today it is centered on Claude Code, with wider agent coverage still forming
SkillCrossroads	Skill and subagent linting	(+)	Audits triggering, least-privilege, token cost, safety, and verifiability	Diagnoses packaging quality rather than runtime agent behavior

Satisfaction was highest when a tool moved state, permissions, or evals out of the model and into something inspectable. The common workaround pattern was to keep a model in the loop for generation but put portability, tool choice, benchmark logic, or command blocking somewhere deterministic. Migration pressure is running from single-vendor agent flows toward mixed stacks where one layer handles sessions, another handles tools, another handles evals, and another handles safety; the competitive question is which layer becomes the default control plane.

5. What People Are Building¶

Project	Who built it	What it does	Problem it solves	Stack	Stage	Links
Skillsync	cat-whisperer	Moves full AI chat sessions across coding agents and makes them searchable	Users get locked into one agent's private transcript format	Rust `txcript` engine, desktop app, CLI, MCP	Beta	post, site, repo
Aclif	chris_marino	Gives agents one CLI grammar across SaaS providers and canonical resource names	Models choose the wrong tool or hold too much provider-specific complexity	TypeScript/oclif, JSON schemas, provider manifests	Beta	post, site, repo
AutoBot	demeyer1	Runs long-horizon AI work with native voice steering, memory, and evidence tracking	Long-running agent work is hard to supervise without babysitting the terminal	Shell harness, voice control, encrypted-disk memory, local ledger	Alpha	post, repo
MCPJam	prathmeshmcp	Tests MCP servers and agent-facing apps across clients before release	Teams cannot see whether external agents actually reach the right result	TypeScript web app, inspector CLI, CI/CD evals, open-source core	Shipped	post, site, repo
Errand	shiqimei	Gives each AI teammate a persistent cloud computer and conversation	Agents lose continuity when sessions or laptops end	macOS app, Runta cloud computer, provider-configured models	Beta	post, site
Insurance Agent Benchmark	jsc39	Measures end-to-end insurance document work on 166 real-world cases	Existing insurance benchmarks do not capture full workflow completion	Cooper document harness, benchmark corpus, model-vs-system comparisons	Beta	post, site
Compute:Arena	prabod	Collects signed local-model benchmark results from community hardware	Local AI performance is hard to compare across chips, quants, and setups	Benchmark harness, signed reports, public leaderboard	Shipped	post, site
GuardRail	promptandbuild	Blocks destructive agent commands and writes audit logs before execution	Teams need pre-execution safety for git, SQL, secrets, and filesystem actions	Bash, jq, openssl, Claude Code hooks, npm package	Beta	post, repo
Skillbay	skeptrune	Marketplace for reusable AI skills curated by a human	Users want reusable agent workflows for tasks outside their expertise	Web marketplace, markdown skills, seller and request flows	Beta	post, site

Skillsync, Aclif, MCPJam, and GuardRail looked like different slices of the same emerging stack: one layer to move sessions, one to standardize tools, one to test outcomes, and one to block dangerous actions. That pattern matters because it shows how much buyer pain now sits outside the model itself.

Errand and AutoBot pushed the same idea into longer-running work. Instead of treating the chat window as the unit of work, both products assume the agent needs durable context, memory, and a place to keep going after the operator walks away.

The evaluation products were more concrete than many earlier benchmark launches. Insurance Agent Benchmark and Compute:Arena both tried to pin performance to a real operating surface - messy documents in one case, specific chips and quants in the other. Skillbay was the outlier: it shows real interest in packaging expertise as reusable agent artifacts, but the HN reaction also shows how quickly that category runs into moat and trust questions.

6. New and Notable¶

Supply-chain risk reached the agent layer¶

Plugin4Shell – Zero Click RCE Vulnerability found in top four coding agents (4 points, 2 comments) mattered because it attacked the layer underneath prompting or model behavior: the marketplace and update path. Air Security's post says a benign plugin can later turn malicious and ride background auto-update into zero-click remote code execution even when teams believe SHA pinning protects them. That makes "review the plugin and pin the commit" look less final than many teams assumed.

AI-assisted rewrites are getting judged as architecture work, not novelty demos¶

Migrating the GitHub Copilot Runtime to Rust, Using Copilot (15 points, 7 comments) was submitted three times on September 17 for a combined 24 points and 7 comments. GitHub's blog post says the runtime behind Copilot CLI, Copilot app, and Copilot SDK was rewritten into more than 800,000 lines of production Rust across 128 pull requests, with agents writing most of the code and the result improving performance by orders of magnitude. HN replies still pressed on code size and whether the original stack choice created avoidable rewrite cost, but the repeated submissions show appetite for concrete architecture stories over abstract "AI makes you faster" claims.

Watermarking is no longer a neutral compliance detail¶

AI model watermarking changes agent behavior (11 points, 0 comments) was notable because it complicates a popular policy assumption: provenance tagging sounds orthogonal to capability, but Lasso's results suggest it can change tool-calling and refusal behavior, especially under prompt injection. That turns watermarking into something that needs red-teaming and agent eval coverage, not just compliance sign-off.

Web operators are already redesigning around agent traffic¶

The agents are coming for the web and the web isn't ready (11 points, 4 comments) was notable because it translated abstract crawler complaints into infrastructure work. The linked Varnish post used public git-hosting traffic to show how agent-driven fetch patterns can make the key space much larger than the content space, then proposed ESI caching so one rendered commit fragment can serve many forks. That is the kind of low-level adaptation that shows AI traffic is no longer hypothetical.

7. Where the Opportunities Are¶

[+++] Portable control planes for heterogeneous agent stacks - Launch HN: Skillsync (YC W26) – AI chat sessions made portable across agents, Show HN: Aclif – Agent CLI framework: one grammar, canonical names across SaaS, Show HN: Open-source AI teammates with their own computer, and Show HN: AutoBot – live voice control for long-running AI work all show direct demand for tools that preserve context, standardize commands, and keep work moving across agents, machines, and time.

[+++] Pre-production evaluation and reliability layers for agent-facing software - Show HN: MCPJam - the first testing & evaluations platform for MCP servers, Insurance Agent Benchmark: 166 real-world cases for evaluating insurance AI, Show HN: Compute:Arena – Community submitted local AI benchmarks, and AI model watermarking changes agent behavior point to a strong need for systems that measure what agents actually do under real clients, real documents, real hardware, and transformed outputs.

[+++] Least-privilege execution and plugin security for coding agents - Plugin4Shell – Zero Click RCE Vulnerability found in top four coding agents, Show HN: GuardRail – 13 guards that stop Claude Code before it pushes to main, Show HN: Linting 216 public Claude Code skills – 69% won't reliably trigger, and Whoever's doing OpenAI's system security is just incompetent in the worst way all converge on the same opportunity: make dangerous capabilities harder to reach by default, and make policy and package quality easier to verify before anything runs.

[++] Persistent workspaces for long-running agents and teams - Show HN: Open-source AI teammates with their own computer, Show HN: AutoBot – live voice control for long-running AI work, and Launch HN: Skillsync (YC W26) – AI chat sessions made portable across agents all show people treating the session less like disposable chat and more like durable work product. The need is strong, but the right surface - local harness, cloud computer, sync layer, or some blend - is still unsettled.

[+] Human-facing agent products with explicit consent and obvious utility - CC is an AI agent for families and groups, Show HN: Craigslist for agent skills, curated by a human, and Show HN: Die With Me – Claude and Codex rate limits as AIM away messages show that the category is active, but the burden of proof is much higher once the agent mediates social relationships or shared household context.

8. Takeaways¶

Attention reconcentrated around safety and applied evaluation. AI safety is mostly a sex cult and Artificial intelligence now beats some of the best human forecasters together captured 46.5% of the day's points and 66.2% of its comments. (source, source)
Safety came back to the front page as a legitimacy and accountability crisis, not a new research consensus. The top story attacked the social roots of AI safety, while the OpenAI incident thread, Karp guidelines story, and Iran deployment story all kept the argument focused on trust and consequences. (source, source, source, source)
Portability across agents is now a real product category. Skillsync, Aclif, Errand, and AutoBot all attacked different forms of lock-in around sessions, tools, workspaces, and long-running tasks. (source, source, source, source)
The harness matters almost as much as the model in evaluation. The forecasting debate, Insurance Agent Benchmark, MCPJam, and watermarking story all pointed to the same lesson: client behavior, document routing, and output transformations can materially change results. (source, source, source, source)
Security talk is moving from model psychology to authority design. Plugin4Shell, GuardRail, the OpenAI air-gap post, and the database-deletion essay all argue that the critical question is what the agent can reach, not whether the model can be scolded into perfect behavior. (source, source, source, source)
Human-facing agent products still face a much higher bar than developer tools. Skillbay, Die With Me, and Google's CC groups launch all show activity in the category, but the HN reactions make clear that social usefulness, consent, and trust are harder sells than another coding harness or benchmark tool. (source, source, source)