Twitter AI Agent - 2026-06-22¶
1. What People Are Talking About¶
1.1 Orchestration moved from harness talk to benchmarked single-endpoint products (🡕)¶
The strongest cluster turned yesterday's harness debate into product claims with concrete public artifacts. At least four strong items supported this theme, and the center of gravity moved from "how to prompt" toward "how to package orchestration, delegation, and verification."
@SakanaAILabs introduced (652 likes, 37 replies, 80,691 views, 349 bookmarks) Sakana Fugu as a full multi-agent orchestration system behind a single model API. The linked Fugu page adds the evidence the launch tweet only hints at: it presents case studies in autonomous ML research, historical document reading, CAD generation, blindfold chess, and stock trading, including an AutoResearch run where Fugu-Ultra executed 123 experiments over about 14 hours on one H100 and finished with the best reported mean BPB.
@SakanaAILabs explained (114 likes, 4 replies, 7,732 views, 30 bookmarks) that Fugu is itself an LLM trained to call other LLMs, including itself recursively, while handling model selection, delegation, verification, and synthesis internally. That architecture mattered because it made the day's central claim explicit: the product being sold is not just the underlying model, but the control layer around a pool of models.
@RoundtableSpace pointed to (81 likes, 11 replies, 51,014 views, 50 bookmarks) Anthropic's Building effective agents guide, which explicitly separates workflows from agents and lays out prompt chaining, routing, parallelization, orchestrator-workers, and evaluator-optimizer patterns. @DanKornas published (34 likes, 5 replies, 1,530 views, 34 bookmarks) AI Development Patterns, a public 24-pattern catalog organized across foundation, development, operations, and experimental tracks, with an explicit harness-engineering lens for feedforward/feedback and computational/inferential controls.
Discussion insight: The most useful pushback was about failure handling, not model quality. One reply to Sakana asked how Fugu maintains verification and consistency on long, high-stakes tasks, while the Anthropic-linked conversation focused on avoiding bad prompting cargo cults by choosing simpler, inspectable patterns first.
Comparison to prior day: June 21 was still dominated by courses, patterns, and enforcement hooks around harnesses. June 22 kept the same subject but advanced it into benchmarked products and official framework write-ups.
1.2 Team-scale agent workspaces became a first-class product category (🡕)¶
A second theme was that the real bottleneck is no longer getting one agent to do one task. It is keeping agents visible, stateful, and governable across teams, sessions, and runtimes. At least five strong items supported this theme.
@huang_chao4969 open-sourced (93 likes, 8 replies, 5,250 views, 75 bookmarks) AgentSpace as a workspace for humans and agents with defined roles, owners, permissions, schedules, and auditability. The public AgentSpace repo makes the design more concrete: it describes AgentRouter support for Claude Code, Codex, OpenClaw, Hermes, and nanobot, plus shared workspaces, approvals, and remote daemon execution.
@goyalshaliniuk argued (28 likes, 10 replies, 427 views, 6 bookmarks) that Claude Code session resets still force developers to re-explain their codebase, and presented Memanto as a local-first fix. Its PyPI page and repo say it provides persistent memory for Claude Code, Cursor, Codex, and 14+ other agents, with remember, recall, and answer primitives, 13 typed memory categories, and a fully local Docker/Ollama path.
@_philschmid announced (39 likes, 9 replies, 2,655 views, 12 bookmarks) Google's Interactions API as generally available. The public Interactions overview says it is now the recommended interface for new Gemini projects and adds server-side state via previous_interaction_id, background execution with background=true, observable execution steps, and a single API surface for both models and agents such as Antigravity and Deep Research.
@DataScienceDojo highlighted (1 like, 170 views, 1 bookmark) DeerFlow 2.0, and the DeerFlow repo describes it as an open-source super-agent harness for sub-agents, memory, sandboxes, and extensible skills. @dew_yashtwt launched Supercode beta (50 likes, 4 replies, 2,522 views, 33 bookmarks) as an open-source terminal SWE agent with full-machine actions, persistent memory, and explicit approval gates; the public repo shows a Bun + Turborepo + Next.js + TypeScript stack for the CLI, terminal web client, dashboard, and shared skills.
Discussion insight: The skepticism here was operational. Supercode replies immediately asked whether the beta was broken, and a reply to the Interactions API launch complained that the official Google .NET SDK still had not caught up. The appetite is clearly there, but the demand is for durable plumbing, not flashy demos.
Comparison to prior day: June 21 leaned toward durable private runtimes and recovery. June 22 pushed the same concern into team workspaces, persistent memory, and unified APIs that let agent work survive beyond one person's terminal.
1.3 Secure action rails kept expanding around payments, skills, and enterprise validation (🡒)¶
The third theme was that agent adoption keeps producing adjacent security and control layers. The day's evidence was less about smarter reasoning than about safer payments, safer capability installs, and more auditable enterprise execution.
@trythreews shipped (214 likes, 48 replies, 19,175 views, 13 bookmarks) a drop-in payment modal for x402 endpoints that handles wallet detection, chain switching, signing, settlement, and browser-side spending caps. @nichxbt added (51 likes, 12 replies, 1,855 views, 8 bookmarks) the bigger frame: Linux Foundation stewardship, major ecosystem backers, more than 165 million transactions, and about 69,000 active agents by late April 2026, while also admitting that much of the current flow is still noisy and speculative.
@probiex007 flagged (1 like, 68 views, 1 bookmark) NVIDIA's SkillSpector, and the repo README says it checks 64 vulnerability patterns across 16 categories while claiming that 26.1% of scanned skills contain vulnerabilities and 5.2% show likely malicious intent. That makes the supply-chain angle on skills much harder to dismiss as abstract.
@GoogleCloudTech reported (63 likes, 2 replies, 6,386 views, 16 bookmarks) that Google used specialized multi-agent systems to migrate large TensorFlow models to JAX. The linked Google Cloud write-up says the system used hierarchical Playbooks, mathematical equivalence checks, and an LLM judge, and delivered a reported 6.4x-8x speedup on complex YouTube model migrations.
Discussion insight: The strongest maturity signal was that the controls are now specific: browser-enforced spend caps, skill scanners, mathematical verification, and quality judges. The most useful note of caution came from the x402 macro thread itself, which said that the current transaction totals still contain substantial speculative noise.
Comparison to prior day: June 21 already showed skill scanning and governance frameworks becoming visible. June 22 made them more operational by tying them to payments UX, install-time scanning, and published enterprise validation loops.
2. What Frustrates People¶
Shared context still disappears too easily¶
Severity: High. @goyalshaliniuk said (28 likes, 10 replies, 427 views, 6 bookmarks) that Claude Code users keep paying a tax every time a session ends: re-explaining the repo, restating decisions, and burning tokens on repeated context. Anthropic's effective agents guide, shared by @RoundtableSpace here (81 likes, 11 replies, 51,014 views, 50 bookmarks), points at the same structural issue from another angle by treating memory, tools, and workflows as first-class building blocks instead of prompt polish. Google's Interactions API docs, announced by @_philschmid here (39 likes, 9 replies, 2,655 views, 12 bookmarks), are effectively a platform-level workaround: server-side state, background jobs, and execution timelines so developers do not have to rebuild context manually. This is worth building for because the pain appears in end-user complaints, framework guidance, and platform design on the same day.
Safe agent execution still needs extra wrappers around the model¶
Severity: High. @trythreews shipped (214 likes, 48 replies, 19,175 views, 13 bookmarks) a payment modal that adds wallet UX, receipts, and browser-side spend caps on top of x402, which is itself evidence that protocol-level payments are not enough for real use. @probiex007 surfaced (1 like, 68 views, 1 bookmark) NVIDIA's SkillSpector, whose README says 26.1% of scanned skills contained vulnerabilities and 5.2% showed likely malicious intent. @GoogleCloudTech showed (63 likes, 2 replies, 6,386 views, 16 bookmarks) the enterprise version of the same problem: their TF-to-JAX migration system needed hierarchical Playbooks, mathematical equivalence checks, and an LLM judge before the output was trusted. The coping pattern is consistent across all three examples: do not trust the raw agent loop by itself.
Team use still breaks when agents live in one person's terminal¶
Severity: Medium. @huang_chao4969 built (93 likes, 8 replies, 5,250 views, 75 bookmarks) AgentSpace around a direct complaint that most agents remain one-person, one-terminal, one-chat-session tools. @dew_yashtwt launched Supercode beta (50 likes, 4 replies, 2,522 views, 33 bookmarks), but the first replies asked if the product was broken, which is a useful reminder that broader machine control increases the operational burden as well as the upside. @DataScienceDojo pointed to (1 like, 170 views, 1 bookmark) DeerFlow 2.0 as a harness with sub-agents, memory, and sandboxes, which again suggests that teams are assembling more infrastructure just to make multi-step agent work manageable. This is worth building for, but the market is already moving from simple chat surfaces toward heavier control planes.
3. What People Wish Existed¶
Shared workspaces that make agents organizational assets¶
The clearest practical need was for a layer that turns private agent setups into shared team infrastructure. @huang_chao4969 described (93 likes, 8 replies, 5,250 views, 75 bookmarks) the problem bluntly: agents are powerful in isolation but break down when real teams need roles, approvals, schedules, and accountability. The public AgentSpace repo now supplies one answer, but the need is broader than one product. Opportunity: direct.
Memory that survives resets without turning into more infrastructure work¶
A second need was for persistent context that does not force developers to stand up a whole retrieval stack. @goyalshaliniuk framed (28 likes, 10 replies, 427 views, 6 bookmarks) the problem as repeated re-explanation after every Claude Code reset, while Memanto's docs position the fix as local-first persistent memory with no vector database to babysit. Google's Interactions API overview, announced by @_philschmid here (39 likes, 9 replies, 2,655 views, 12 bookmarks), points to the same demand from the platform side with server-managed state and background execution. Opportunity: direct.
Safer install, payment, and approval layers around agent actions¶
The feed kept implying the same missing product category: wrappers that make agent actions auditable before, during, and after execution. @probiex007 shared (1 like, 68 views, 1 bookmark) SkillSpector's install-time scanner; @trythreews added (214 likes, 48 replies, 19,175 views, 13 bookmarks) wallet UX, receipts, and spend caps for x402 flows; and @GoogleCloudTech showed (63 likes, 2 replies, 6,386 views, 16 bookmarks) that even an internal migration system needed hard verification loops before the output could be trusted. The need is practical and immediate, though competitive. Opportunity: direct.
4. Tools and Methods in Use¶
| Tool | Category | Sentiment | Strengths | Limitations |
|---|---|---|---|---|
| Sakana Fugu | Orchestration model/API | (+/-) | Single endpoint for model routing, delegation, verification, and synthesis; public case studies across research and coding-adjacent tasks | Replies immediately questioned consistency, cost, and how much is genuinely new versus other wrappers |
| AgentSpace | Shared agent workspace | (+) | Roles, owners, approvals, audit trails, scheduling, and AgentRouter across multiple runtimes | Very new product category; still invites questions about how it differs from adjacent workspace tools |
| Memanto | Memory agent | (+) | Local-first persistent memory, no vector DB, broad coding-agent integrations, typed memory categories | Adds another component to manage and today had more builder interest than broad user validation |
| Interactions API | Agent/model API | (+) | Single interface for models and agents, server-side state, background jobs, observable execution steps | SDK parity is still incomplete; per-interaction settings must still be restated |
| AI Development Patterns | Workflow catalog | (+) | 24-pattern map for adopting harness, security, and development controls | Guidance only; does not enforce compliance by itself |
| DeerFlow 2.0 | Super-agent harness | (+) | Sub-agents, memory, sandboxes, extensible skills, open source under MIT | More setup-heavy than lightweight prompt or chat tools |
| SkillSpector | Security scanner | (+) | 64 patterns across 16 risk categories, CVE lookups, CI-friendly outputs | Focused on install-time skill review, not full runtime governance |
| Supercode | Terminal SWE agent | (+/-) | Full-machine actions, approvals, persistent memory, multi-model support, web + CLI surfaces | Broad permissions and beta reliability concerns raise operational risk |
| x402 | Payment protocol | (+/-) | Machine-payable HTTP endpoints, expanding ecosystem support, usable modal wrappers now appearing | Early activity is still noisy and the wallet UX is hard without extra layers |
| GLM 5.2 | Open-weight model/API distribution | (+) | AWS Marketplace and Akash availability reduced access friction for coding-agent workloads | Most evidence today was about distribution and positioning, not deep production case studies |
The overall satisfaction spectrum was strongest for explicit control layers around the model: workspaces, memory, scanners, playbooks, and stateful APIs. It was more mixed for broad-surface beta agents and anything that claimed autonomy without equally explicit safeguards.
A clear migration pattern kept showing up: people are moving away from one-off prompt craft toward harnesses, shared workspaces, memory layers, and verification loops. Even model distribution news followed that pattern. @CarolGLMs announced (299 likes, 16 replies, 40,310 views, 41 bookmarks) GLM 5.2 on AWS Marketplace, while @akashnet added (43 likes, 2 replies, 3,303 views, 2 bookmarks) Akash availability for the same model, so the open-model conversation today was mostly about where builders can plug these models into existing harnesses.
5. What People Are Building¶
| Project | Who built it | What it does | Problem it solves | Stack | Stage | Links |
|---|---|---|---|---|---|---|
| Sakana Fugu | @SakanaAILabs | Multi-agent orchestration model behind a single API | Hides routing, delegation, and verification for hard multi-step tasks | OpenAI-compatible API, orchestrated LLM pool, recursive model calling | Shipped | launch, architecture, site |
| AgentSpace | @huang_chao4969 | Shared workspace where humans and agents work with roles, approvals, and audit trails | Turns private agent setups into visible organizational assets | Node.js 24, PostgreSQL 16, AgentRouter, Claude Code/Codex/OpenClaw/Hermes/nanobot | Shipped | tweet, repo |
| Supercode | @dew_yashtwt | Terminal SWE agent with full-machine actions and a matching web client | Gives a single agent access to files, commands, apps, and web actions with approvals | Bun, Turborepo, TypeScript, Next.js, Anthropic/OpenRouter/Google providers | Beta | tweet, site, repo |
| Memanto | @goyalshaliniuk | Companion memory agent for coding agents | Preserves context across session resets without a separate vector DB stack | Python CLI, local Docker/Ollama path, Moorcheh search engine, MIT | Shipped | tweet, PyPI, repo |
| DeerFlow 2.0 | ByteDance | Open-source super-agent harness with sub-agents, memory, sandboxes, and skills | Makes long-horizon research or coding work easier to structure and isolate | Python 3.12+, Node.js 22+, extensible skills, sandbox mode, memory | Shipped | tweet, repo |
| SkillSpector | NVIDIA | Skill security scanner | Checks agent skills for vulnerabilities or malicious patterns before install | Python 3.12+, static analysis, optional LLM review, OSV lookups, Apache 2.0 | Shipped | tweet, repo |
| x402 payment modal | @trythreews | Browser modal for paying x402 endpoints | Adds wallet UX, receipts, spend caps, and retries to machine-payable HTTP endpoints | x402, Solana helpers, browser modal, USDC/$THREE | Shipped | tweet |
AgentSpace, Supercode, and Memanto attacked the same operational problem from different sides: visibility, action surface, and continuity. One makes agents shareable and governable, one makes them capable of whole-machine work, and one keeps them from forgetting everything between sessions.
Sakana Fugu and DeerFlow represented the stronger harness-side builds. Fugu packages orchestration as a model product, while DeerFlow exposes the more inspectable open-source version of the same instinct: split the job across sub-agents, preserve memory, and contain risky execution in sandboxes.
SkillSpector and the x402 modal were the control-layer counterparts. One treats installable skills as a supply-chain surface to scan before use, and the other treats agent payments as something that needs explicit UX and boundaries, not just a protocol spec. Google's TF-to-JAX migration system was the enterprise internal analogue: another example where the build itself is only acceptable once the verification loop is part of the product.
6. New and Notable¶
Google's Interactions API became the default path for new Gemini agent projects¶
@_philschmid announced (39 likes, 9 replies, 2,655 views, 12 bookmarks) that the Interactions API is now generally available, and the public overview says it is the recommended interface for all new Gemini projects. That matters because it collapses model calls and agent calls into one surface while adding server-side state, background jobs, and observable execution steps.
A concrete enterprise migration win put multi-agent validation in public¶
@GoogleCloudTech reported (63 likes, 2 replies, 6,386 views, 16 bookmarks) a specialized multi-agent system for migrating production TensorFlow models to JAX, and the linked Google Cloud write-up says it delivered a 6.4x-8x speedup on complex YouTube models. The notable part was not just the speedup, but the validation stack: hierarchical Playbooks, mathematical equivalence checks, and an LLM judge.
7. Where the Opportunities Are¶
[+++] Shared agent control planes for teams — AgentSpace, Memanto, Interactions API, DeerFlow, and Supercode all point to the same gap: agents need shared context, resumable work, approvals, and runtime portability if they are going to operate as part of a real team instead of as one person's private tool. This is strong because the need shows up in direct complaints, new products, and platform primitives at once.
[++] Security and validation wrappers for agent actions — SkillSpector's scanner, the x402 modal's spend caps and receipts, and Google's mathematically verified migration system all show demand for products that make agent behavior inspectable and bounded. This is moderate because the need is obvious, but several builders are already converging on it from different angles.
[+] Open-model distribution plus orchestration adapters — Sakana Fugu, GLM 5.2 on AWS Marketplace, and GLM 5.2 on Akash all point to growing demand for alternative model access paths that fit existing coding harnesses and cloud workflows. This is emerging because the distribution signal is real, but differentiation remains crowded and still leans more on availability than on proven downstream outcomes.
8. Takeaways¶
- Orchestration is becoming a product surface, not just an implementation detail. Sakana Fugu was the clearest example, and Anthropic's public framework gave the same idea a simpler architectural vocabulary. (source)
- The next agent battleground is team operations, not solo prompting. AgentSpace, Memanto, Interactions API, DeerFlow, and Supercode all focused on shared context, resumability, approvals, or sandboxed execution rather than on a single bigger model. (source)
- Safe agent action still needs explicit wrappers around the loop. The x402 modal added spend caps and receipts, SkillSpector treated skills as a risky install surface, and Google's migration system required hard verification before shipping output. (source)
- Open-model momentum showed up mostly as distribution, not yet as rich deployment evidence. GLM 5.2's strongest signals today were AWS Marketplace and Akash availability for coding-agent workloads, not deeply documented production stories. (source)
- The conversation keeps moving away from prompts and toward harnesses. AI Development Patterns, Anthropic's workflow taxonomy, and the persistent-memory products all treated the durable system around the model as the real source of leverage. (source)