Twitter AI Agent - 2026-06-24¶
1. What People Are Talking About¶
1.1 Persistent teammate agents moved from demos to everyday work surfaces (🡕)¶
The strongest conversation was no longer about a single chat window. It was about agents becoming persistent coworkers that live inside shared work surfaces, keep context, and keep running after the human steps away. At least four retained items supported this theme.
@karpathy argued (7,076 likes, 389 replies, 917,172 views, 3,866 bookmarks) that Claude has entered a “third major redesign” of LLM UI, from website to desktop app to a self-contained, persistent, asynchronous entity with org-wide tools and context. The quoted Claude Tag launch makes the claim concrete: Claude joins Slack as a scoped teammate, remembers channel context, can work asynchronously, and ships with spend limits and audit logs.
@AlexFinn reported (248 likes, 17 replies, 18,688 views, 223 bookmarks) that Hermes now defaults to background agents, adds a skills hub, exposes live subagent panes, and lets users message the system through iMessage or Telegram. The notable part was not any single feature, but the product assumption underneath it: the agent is expected to keep running while the user changes device or context.
@cryptoleek flagged (77 views) that PI WEB keeps Pi Coding Agent sessions alive across browser closes and server restarts. The public repo says the browser UI sits in front of a separate session daemon and Pi SDK, so long-running workspaces and sessions survive UI restarts instead of dying with the tab.
Discussion insight: The most useful pushback was about whether this is genuinely new or just better packaging. Replies to Karpathy immediately asked what was different from older “tag an agent” setups, and the answer centered on enterprise-grade deployment, scoped tools, memory, and multiplayer continuity rather than basic chat invocation.
Comparison to prior day: June 22 already showed team workspaces, memory layers, and unified agent APIs. June 24 pushed that one step further into shared, always-on team surfaces that behave more like coworkers than like tabs.
1.2 Harness and loop engineering kept displacing prompt craft (🡕)¶
A second theme was that the operational skill of the day was not “prompt better.” It was designing loops, externalizing state, packaging workflows into skills, and adding guardrails that survive across runs. The strongest evidence came from both high-engagement explanation threads and lower-engagement but concrete practitioner playbooks.
@akshay_pachaar wrote (797 likes, 37 replies, 103,374 views, 1,337 bookmarks) that loop engineering matters because the human is otherwise still the scheduler and checker inside the loop. His thread named the missing pieces directly: triggers, maker/checker separation, disk-backed state, explicit stop conditions, temporal memory via Graphiti, and harness observability via Opik.
@swadeshkumar_ argued (6 likes, 85 views) that Claude Code works best when the repository itself is shaped for the agent: short CLAUDE.md files, reusable .claude/skills/, hooks for mandatory checks, and documentation the agent can navigate instead of having everything stuffed into prompts. That was a small post, but it translated the day’s abstract loop talk into a concrete repository layout.
@charliejhills compiled (24 likes, 7 replies, 3,161 views) a 17-link Claude Skills resource pack, including Anthropic’s agent skills best practices. Separately, the agent-skills repo describes skills as production-grade workflow packages with slash commands spanning spec, plan, build, test, review, and ship.
Discussion insight: The recurring nuance was that state has to live outside the model. One reply to Akshay’s thread said the important thing is not one long-running agent versus many agents, but preserving objective, tasks, status, evidence, and exit criteria outside the context window.
Comparison to prior day: June 22 was still heavy on orchestration products and pattern catalogs. June 24 shifted the center of gravity toward the engineering layer around the model: loops, files, hooks, and skills.
1.3 Memory, skill vetting, and least-privilege control became explicit infrastructure (🡕)¶
The feed also treated several formerly hand-wavy agent concerns as concrete infrastructure categories: memory systems with measurable modules, install-time scanning for third-party skills, and authorization models that derive scope from the specific task instead of from a permanently overpowered assistant.
@dair_ai summarized (128 likes, 14 replies, 7,612 views, 145 bookmarks) a long-term memory paper by saying agent memory is now a data-management layer with storage, retrieval, update, consolidation, and lifecycle governance. The claim mattered because it rejected end-to-end task success as the only metric and instead pushed memory toward module-level evaluation.
@KentonVarda argued (99 likes, 13 replies, 6,818 views, 78 bookmarks) that explicit per-agent permission setup is the wrong safety model because it does not scale. His alternative was capability-based security, where each new task gets its own agent and its authority is inferred from the specific resources the human provided, with all agent authority remaining a subset of a human’s authority.
@JafarNajafov highlighted (9 likes, 718 views) a cluster of fast-rising repos: headroom for 60-95% token reduction, agent-skills for packaged engineering workflows, and SkillSpector for pre-install skill scanning. SkillSpector’s public README says 26.1% of scanned skills contained vulnerabilities and 5.2% showed likely malicious intent, making skill installation look more like supply-chain security than casual prompt sharing.
@NostaIgicGareth pointed to (494 views) agentmemory, whose public repo positions itself as persistent memory for Claude Code, GitHub Copilot CLI, Cursor, Gemini CLI, Codex CLI, Hermes, OpenClaw, pi, and MCP clients. The README’s own badges claim 95.2% retrieval R@5 and 92% fewer tokens, which aligned with the day’s broader push to make memory both cheaper and more durable.
Discussion insight: Replies around both the memory and auth posts converged on the same principle: agents should not be trusted because they are clever. They should be trusted only when memory, permissions, and verification are made explicit.
Comparison to prior day: June 22 already surfaced memory and governance as adjacent concerns. June 24 made them more operational with module-level memory framing, install-time scanners, and explicit least-privilege arguments.
1.4 Agent products kept expanding into sales, cloud operations, payments, and research (🡕)¶
The last major theme was breadth. The day’s builders were not only shipping coding helpers. They were applying the same agent patterns to outbound sales, cloud infrastructure, machine payments, and packaged research services.
@austinh___ announced (219 likes, 56 replies, 240,357 views, 144 bookmarks) a chat product for outbound sellers that works across 40+ data sources. The strongest evidence in the post was operational rather than conceptual: 57,548 queries in its first few weeks of beta and 45% week-over-week growth.
@mvcinvesting reported (467 likes, 21 replies, 30,593 views, 33 bookmarks) that Nebius AI Cloud 3.6 adds Echo, a natural-language agent for infrastructure control, alongside KMS, customer-managed encryption keys, Workload Identity Federation, and large storage upgrades. Public release coverage of Nebius AI Cloud 3.6 matches that framing.
@ampersend_ai explained (7 likes, 386 views, 3 bookmarks) a pay-per-intelligence routing layer on Amazon Bedrock AgentCore Payments. The companion AWS post describes a two-hop payment flow where the agent pays Ampersend and Ampersend settles with the upstream model provider, with x402, budgets, and auditability handled underneath.
@Capafyai introduced (125 views) Deep Research Pro as a sellable skill agent that turns questions into citation-backed reports. The public Capafy skills repo and deep-research-pro repo both make the pattern explicit: research workflows are being packaged as closed-source or reusable services, not just kept as private prompts.
Discussion insight: The interesting commonality was that these products all wrapped extra infrastructure around the model: data connectors for sales, governance for cloud control, managed payments for autonomous calls, and marketplace/IP rails for research skills.
Comparison to prior day: June 22 emphasized orchestration systems and team workspaces. June 24 showed the same design logic being commercialized across adjacent business functions.
2. What Frustrates People¶
Context still disappears unless teams build memory outside the model¶
Severity: High. @akshay_pachaar said (797 likes, 37 replies, 103,374 views, 1,337 bookmarks) that state has to live on disk, not in context, because the model forgets between runs. @dair_ai framed (128 likes, 14 replies, 7,612 views, 145 bookmarks) the same problem more structurally: memory now needs storage, retrieval, updates, consolidation, and governance, but most teams still evaluate it as a black box. @NostaIgicGareth surfaced (494 views) agentmemory precisely because people are tired of re-explaining the codebase every session. The coping pattern is clear: persistent memory layers, compressed context, and file-backed state. This is worth building for because multiple unrelated posts converged on the same pain.
Authorization and safety break when agents start over-privileged¶
Severity: High. @KentonVarda argued (99 likes, 13 replies, 6,818 views, 78 bookmarks) that having to pre-configure each agent’s permissions does not scale and encourages assistants that start with access to too much. A reply to @karpathy added (7,076 likes, 389 replies, 917,172 views, 3,866 bookmarks) the operational version of the same concern: if Claude is a teammate, it also needs IAM, audit trails, and a clean way to stop it before it does something expensive. The SkillSpector repo, surfaced by @JafarNajafov (9 likes, 718 views), adds another safety complaint at install time by claiming 26.1% of scanned skills contain vulnerabilities and 5.2% show likely malicious intent. Today’s workaround is to add scanners, scoped capabilities, and explicit human-derived authority instead of trusting broad default access.
Production agents still need a lot of plumbing around money, cloud control, and observability¶
Severity: High. @ampersend_ai described (7 likes, 386 views, 3 bookmarks) a world where every paid model or data endpoint otherwise forces developers to rebuild wallet custody, x402 support, spend limits, and billing integrations. @mvcinvesting showed (467 likes, 21 replies, 30,593 views, 33 bookmarks) that even cloud control now ships with KMS, Workload Identity Federation, and budgets because natural-language control by itself is not enough. @akshay_pachaar also said observability has to catch harness drift after models or prompts change, not just confirm that one run passed. This looks worth building for because the same extra-plumbing problem appeared in payment, cloud, and loop-management posts on the same day.
3. What People Wish Existed¶
Shared, persistent agent identities that survive handoffs¶
The clearest practical need was for agents that behave like durable teammates instead of one-session assistants. @karpathy described (7,076 likes, 389 replies, 917,172 views, 3,866 bookmarks) the desired state as an org-wide, asynchronous entity with tools and memory, while Anthropic’s Claude Tag announcement adds channel scoping, ambient updates, and spend controls. @cryptoleek showed (77 views) the same demand from the coding side with PI WEB, where sessions survive browser closes and machine changes. Opportunity: direct.
Repo-native workflow packaging instead of repeated prompt babysitting¶
People also want reusable workflow structure to beat prompt drift. @swadeshkumar_ argued (6 likes, 85 views) for short CLAUDE.md files, local context files, hooks, and skills, while @charliejhills collected (24 likes, 7 replies, 3,161 views) documentation and examples for exactly that workflow. The agent-skills repo makes the same demand explicit by mapping skills to spec, plan, build, test, review, and ship phases. Opportunity: direct.
Memory and compression layers that preserve signal without exploding token cost¶
A second strong need was for systems that keep the useful parts of prior work while cutting down context bloat. @dair_ai framed memory as a governed data system, @NostaIgicGareth pointed to persistent memory via agentmemory, and @JafarNajafov surfaced headroom as a compression layer claiming 60-95% fewer tokens. The need is practical because the products are being sold as fixes for current session limits, not as abstract research. Opportunity: direct.
Managed infrastructure for payments, permissions, and trust around autonomous action¶
The feed also implied a need for agent infrastructure that feels as boring and reliable as auth or billing in web apps. @ampersend_ai argued (7 likes, 386 views, 3 bookmarks) that developers should not have to build wallet custody and spend controls before shipping agent logic, while @KentonVarda argued for capability-scoped task authority instead of broad standing permissions. @JafarNajafov added the install-time version of the same need by surfacing SkillSpector. Opportunity: direct.
4. Tools and Methods in Use¶
| Tool | Category | Sentiment | Strengths | Limitations |
|---|---|---|---|---|
| Claude Tag | Team agent surface | (+) | Shared Slack-native context, async work, channel-scoped memory, admin controls | Beta product and still invites questions about how much is truly new versus existing taggable agents |
| Hermes Agent | Personal agent runtime | (+/-) | Background agents, multi-profile workflows, skills hub, mobile messaging surfaces | Replies still asked about setup friction, VPS needs, and gaps between channels |
| Graphiti | Memory graph | (+) | Temporal context graphs, provenance, hybrid retrieval, evolving facts | Adds another system layer to operate |
| Opik | Observability/evals | (+) | Trace capture, evaluation, prompt and tool optimization, production monitoring | Useful only if teams instrument their harnesses and keep reviewing traces |
| agentmemory | Persistent memory layer | (+) | Cross-client memory, compression, MCP support, file-backed recall | Early open-source momentum is strong, but broad production evidence is still limited |
| headroom | Context compression | (+) | 60-95% token reduction claims, library/proxy/MCP modes, cross-agent memory options | Compression adds another preprocessing layer that teams must trust and tune |
| agent-skills | Workflow packaging | (+) | Development-lifecycle skills, slash commands, reusable quality gates | Competitive space, and quality still depends on the skill author |
| SkillSpector | Skill security scanner | (+) | Pre-install scanning, 68 patterns across 17 categories, SARIF/JSON output | Focuses on install-time review; runtime abuse still needs separate controls |
| Nebius Echo / AI Cloud 3.6 | Cloud operations | (+/-) | Natural-language infra control plus KMS, CMEK, WIF, storage, and search upgrades | More capability also means more governance burden before touching production |
| PI WEB | Agent control plane | (+) | Persistent sessions, split-process daemon, multi-device supervision, worktree support | Pi-specific and still centered on one agent ecosystem |
| Deep Research Pro / Capafy | Research skill + marketplace | (+/-) | Cited reports, multi-source fetch, monetizable skill distribution, IP protection | Marketplace quality and repeatability remain open questions |
| Ampersend + AgentCore Payments | Agent payments | (+) | Two-hop pay-per-intelligence flow, budgets, x402, unified provider access | Depends on payment rails and governance layers that many teams do not yet have |
The overall satisfaction spectrum favored tools that make agents more inspectable: memory layers, skills, scanners, traces, and scoped control planes. The weaker sentiment appeared when a product asked users to trust broad autonomy without equally strong guardrails.
The clearest workaround pattern was to move critical state and policy out of the model. Teams are storing memory on disk or in graphs, compressing context before it hits the model, packaging repeatable work into skills, and inserting scanners or spend controls before agents touch real systems.
Migration pressure also showed up in the tool mix. Instead of one premium assistant doing everything, the feed kept pointing toward layered stacks: a teamwork surface such as Claude Tag, a memory layer such as agentmemory or Graphiti, a workflow layer such as agent-skills, and a safety layer such as SkillSpector or AgentCore Payments.
5. What People Are Building¶
| Project | Who built it | What it does | Problem it solves | Stack | Stage | Links |
|---|---|---|---|---|---|---|
| Claude Tag | @claudeai / @karpathy | Shared Slack-native Claude teammate with scoped channels, tools, and async tasks | Teams need persistent, multiplayer agent work instead of isolated chats | Slack, org-scoped tools, channel memory, admin controls | Beta | tweet, announcement |
| Claude for outbound sellers | @austinh___ | GTM assistant that works across 40+ data sources through chat | Sellers want agentic outbound help without technical setup | Proprietary app, 40+ data sources | Beta | tweet |
| agentmemory | @NostaIgicGareth / @ghumare64 | Persistent memory for coding agents across multiple clients | Re-explaining the codebase every session wastes tokens and attention | iii engine, MCP, hooks, multi-client integrations | Shipped | tweet, repo |
| PI WEB | @cryptoleek / jmfederico | Browser control plane for long-running Pi Coding Agent sessions | Sessions usually die with the terminal, browser tab, or machine handoff | Fastify Web/API process, session daemon, Pi SDK, git worktrees | Beta | tweet, repo |
| SkillSpector | @JafarNajafov surfacing NVIDIA | Scanner for AI agent skills before installation | Teams need to vet third-party skills for vulnerabilities and malicious patterns | Python, OSV.dev, SARIF/JSON output, optional LLM analysis | Shipped | tweet, repo |
| Deep Research Pro on Capafy | @Capafyai | Research skill that turns questions into citation-backed reports and can be sold as a service | Generic assistants often stop at shallow summaries, and creators want to monetize stronger workflows | Skill agent, multi-source search/fetch, Capafy marketplace | Beta | tweet, Capafy skills, deep-research-pro |
| Ampersend | @ampersend_ai | Pay-per-intelligence routing layer between agents and model providers | Agent builders do not want to reimplement wallet custody, budgets, and payment orchestration for every provider | Amazon Bedrock AgentCore Payments, x402, USDC on Base, provider routing | Shipped | tweet, AWS post |
Claude Tag was the day’s clearest signal that agent products are moving from personal productivity into shared organizational workflow. The important part was not just “Claude in Slack,” but the scoped identity, shared visibility, async follow-up, and Anthropic’s claim that this pattern already drives most of its own product-team code output.
agentmemory and PI WEB attacked the same persistent-work problem from different sides. agentmemory tries to preserve the right context across many coding clients, while PI WEB keeps the execution environment and session daemon alive so the work itself can outlast the front-end session.
SkillSpector, Deep Research Pro, and Ampersend point to three adjacent build patterns. One packages expertise as a reusable skill, one sells that skill as a service with IP protection, and one adds the billing and budget layer so autonomous agents can buy intelligence without hand-built payment plumbing each time.
6. New and Notable¶
Claude Tag made the “agent as coworker” interface concrete¶
@karpathy turned (7,076 likes, 389 replies, 917,172 views, 3,866 bookmarks) a product launch into the day’s defining framing shift: the agent as a self-contained, persistent, asynchronous teammate. The accompanying Anthropic announcement added the public specifics that made the claim more than marketing: channel-scoped identities, shared visibility, ambient follow-up, spend limits, and logs.
Install-time skill scanning became a visible new category¶
@JafarNajafov surfaced (9 likes, 718 views) SkillSpector, and the repo’s claim that 26.1% of scanned skills contain vulnerabilities and 5.2% show likely malicious intent gave the feed a concrete supply-chain statistic for agent tooling. That is notable because the conversation is moving from “share your prompt pack” toward “scan the thing before you install it.”
AWS-backed agent payments moved from concept to implementation detail¶
@ampersend_ai showed (7 likes, 386 views, 3 bookmarks) that pay-per-intelligence infrastructure is becoming concrete enough for AWS case studies. The companion AWS write-up lays out the payment manager, two-hop routing flow, budgets, and x402 settlement path in enough detail to treat agent commerce as real infrastructure, not just as a crypto-adjacent idea.
7. Where the Opportunities Are¶
[+++] Persistent agent control planes — Evidence spans section 1.1, section 2, and section 5: Claude Tag, Hermes background agents, PI WEB, and agentmemory all point to the same need for sessions, memory, and handoffs that survive context loss and device changes.
[+++] Authorization, scanning, and spend-governance layers — Kenton Varda’s least-privilege argument, SkillSpector’s vulnerability claims, Nebius’s governance upgrades, and Ampersend’s payment budgets all point to the same missing stack: trustworthy control over what agents may access, buy, and execute.
[++] Workflow packaging for coding agents — The day’s skills posts, repo-structure advice, and rising repos show demand for reusable engineering workflows that reduce prompt babysitting. The space is already getting crowded, but the need is direct and current.
[+] Vertical agent applications with real operating data — The outbound-sales product’s reported 57,548 beta queries and 45% weekly growth suggest room for domain-specific agents that sit on top of rich data connectors. The opportunity is emerging, but today’s evidence came from fewer products than the infrastructure themes did.
8. Takeaways¶
- The center of gravity moved toward persistent, shared agent work rather than isolated chat sessions. Karpathy’s framing plus Anthropic’s public Claude Tag details made that shift explicit. (source)
- Loop engineering is now about external state, checkers, and traces, not just better instructions. Akshay Pachaar’s thread tied the workflow to Graphiti for memory and Opik for observability, while practitioner posts translated it into repo structure and skills. (source)
- Memory, permissions, and skill safety are becoming first-class infrastructure categories. The memory paper thread, capability-based authorization argument, agentmemory repo, and SkillSpector scan claims all reinforced the same pattern. (source)
- Agent commercialization is broadening beyond coding into cloud operations, payments, research, and sales. Nebius, Ampersend, Capafy, and the outbound-sales product all showed concrete business surfaces wrapping infrastructure around the model. (source)