Twitter AI Agent - 2026-06-20¶
1. What People Are Talking About¶
1.1 Loop engineering broadened into harness and context-layer work (🡒)¶
The loudest technical cluster was still about loops, but the June 20 discussion was more specific about the engineering wrapped around them: harness design, evals, persistent context, and control systems. At least four strong items supported this theme.
@0xMovez shared (274 likes, 25 replies, 56,401 views, 545 bookmarks) Boris Cherny's line that loops are as large a step as the move from source code to agents, and that roughly 30% of his own code is already fully written by loops. The distinctive part was not just the slogan; replies immediately treated the loop-engineering article and dynamic-workflow discussion as the next reading list for serious builders.
@Vtrivedy10 argued (94 likes, 7 replies, 8,216 views, 147 bookmarks) that good harnesses are built around model-harness-task fit, evals, trace inspection, and clean per-subtask context windows. In replies, he added that progress after the easy wins is mostly driven by the quality of the reward signals and evals teams can extract from traces.
@GergelyOrosz observed (110 likes, 25 replies, 10,297 views, 33 bookmarks) a surge of VC-funded and bootstrapped efforts trying to build a "context layer for engineering teams." The useful nuance came from replies: the valuable version is not one-shot retrieval over docs, but persistent memory that compounds across sessions and systems.
@rohit_tiwari pointed to (37 likes, 4 replies, 1,261 views, 36 bookmarks) the free Learn Harness Engineering course, framing harness work as environment design, state management, verification, and control systems rather than prompt tweaking.
Discussion insight: The replies repeatedly drew the same line: retrieval is not memory, vibes are not evals, and loops only become reliable when someone is explicitly managing context, traces, and failure analysis.
Comparison to prior day: June 19 made "loop engineering" a major phrase in the feed. June 20 extended that into public courseware and a clearer market category around context layers for engineering teams.
1.2 Agent definitions converged on folders, markdown, and opt-in components (🡕)¶
A second theme was a growing consensus that the cleanest way to describe an agent is as files in a directory, not as a hidden runtime bundle. At least four items supported this convergence.
@rauchg argued (546 likes, 60 replies, 33,697 views, 240 bookmarks) that the next hot programming language is markdown, showing a minimal eve agent as an agent/ folder with instructions.md and a skills/ directory. In replies, he added that eve exposes a typed useEveAgent() hook for custom UIs, which makes the folder layout more than a slogan.
@hwchase17 highlighted (49 likes, 6 replies, 3,965 views, 37 bookmarks) Leve, and the public Leve repo describes it as a filesystem-first durable agent framework built on LangGraph. The repo adds the implementation detail the tweet only hints at: approvals, subagents, connections, schedules, evals, sandboxed compute, and per-caller security are all wired into the directory-based definition.
@alex_prompter compiled (46 likes, 9 replies, 3,306 views, 89 bookmarks) a list of highly starred repos whose main payload is markdown, skills, or prompts rather than runtime code. Even with some promotional framing, the post captured a real distribution pattern: behavior layers are now being shared as repos in their own right.
@sudoingX praised (76 likes, 12 replies, 5,785 views, 45 bookmarks) Hermes Agent's new blank-slate setup mode, where the default agent starts with a provider, a model, file operations, and a terminal, and everything else is added back intentionally. Replies made the appeal explicit: users want the persistent shell or web fetch first, not a giant preloaded surface they may never use.
Discussion insight: The convergence was not just on markdown. It was on inspectability: people want the agent definition to live in plain files, and they increasingly prefer opt-in tools over all-in-one defaults.
Comparison to prior day: June 19 treated coding-agent workspaces as operating systems. June 20 narrowed the winning abstraction further, toward the file tree, the skill folder, and the minimal default install.
1.3 General-purpose agents were judged by whether they could finish real workflows (🡕)¶
A third theme was that the strongest workflow posts were no longer about code generation alone. They were about whether the agent could move across services, devices, and channels and actually complete the job.
@AlexFinn said (198 likes, 34 replies, 11,478 views, 157 bookmarks) that Codex wrote a newsletter landing page, pushed it to GitHub, created a Vercel project, connected the repo, chose the right domain, and got the page live in about five minutes. The important follow-up came from replies: several readers said the more interesting unlock is giving the agent boring desktop and browser tasks first, not saving it only for engineering.
@IBuzovskyi shipped (74 likes, 3 replies, 10,678 views, 58 bookmarks) Hermes Agent v0.17 with iMessage support, background subagents, image editing, automation blueprints, a profile builder, skill-browser security scans, atomic memory operations, and dashboard hardening. The thread mattered because it read like a real release note, not a vague product teaser.
@rabbit_hmi rolled out (73 likes, 13 replies, 9,083 views, 22 bookmarks) rabbitOS 2.2 with Claude Code sessions on r1, terminal mode, and voice control for staying connected to active work. The tweet was notable less for the hardware image than for the claim that coding-agent sessions are now expected to follow users onto ambient devices.
Discussion insight: The most consistent skepticism was about fragmentation, not capability. Replies to Alex Finn's post asked how users are supposed to connect Claude Code, Codex, and other top-tier tools instead of constantly pivoting between them.
Comparison to prior day: June 19 treated the agent workspace itself as the product. June 20 pushed that workspace outward into browsers, messaging channels, phones, and voice-controlled hardware.
1.4 Trading and agent-commerce posts stayed active, but the best evidence came from open frameworks and explicit stacks (🡒)¶
The feed stayed loud on trading and commerce, but the strongest evidence was not "easy profit" rhetoric. It came from research frameworks, concrete stack diagrams, and launch mechanics.
@quantscience_ shared (34 likes, 4 replies, 3,124 views, 52 bookmarks) TradingAgents, and the public TradingAgents repo describes it as a multi-agent trading framework with analyst, researcher, trader, risk-management, and portfolio-manager roles. The paper's arXiv abstract says the framework beat baselines on cumulative return, Sharpe ratio, and maximum drawdown across its sampled experiments.
@cyrilXBT assembled (101 likes, 17 replies, 5,783 views, 85 bookmarks) a six-link open stack for an automated trader, using Lightweight Charts, CCXT, Hyperliquid's Python SDK, and Tavily among the components. The most valuable part of the thread was actually the pushback in replies, where one reader said Binance rate-limit problems alone consumed three days of debugging.
@slash1sol announced (98 likes, 28 replies, 5,222 views, 72 bookmarks) Tiny Place as a live social economy for agents on Solana, where agents can discover bounties, negotiate fees, and settle in USDC via x402 while keeping custody in Phantom wallets. The evidence is still launch-stage, but it is a concrete attempt to turn agents into market participants rather than chat surfaces.
@Cryptic_Web3 reported (211 likes, 31 replies, 20,748 views) that aipaywithcrypto raised $10 million to build AI-powered payments infrastructure, a marketplace, and native payment integrations for autonomous agents.
Discussion insight: The strongest reality check came from replies, not detractors. People immediately asked whether any of these stacks are actually profitable in live conditions and how much "zero-cost" automation survives first contact with rate limits and debugging.
Comparison to prior day: June 19's commerce thread focused on approvals, signing, and proof. June 20 shifted toward open trading frameworks, agent marketplaces, and financing for payment rails.
2. What Frustrates People¶
Reliable harnesses still depend on trace work, context discipline, and human review¶
Severity: High. @Vtrivedy10 said (94 likes, 7 replies, 8,216 views, 147 bookmarks) that teams leave performance on the table when they "vibe" a harness instead of driving it with evals, pass-rate data, token consumption, and traces. @GergelyOrosz added (110 likes, 25 replies, 10,297 views, 33 bookmarks) a second version of the same problem: teams want Claude Code to know what lives in email, Jira, tickets, and calls, but the hard part is building a persistent context layer instead of reloading the window every session. @rohit_tiwari framed (37 likes, 4 replies, 1,261 views, 36 bookmarks) the workaround as explicit environment design, state management, verification, and control systems. This is worth building for because the pain shows up in practitioner notes, market demand, and public training material on the same day.
All-in-one agent setups still feel bloated or disconnected¶
Severity: Medium. @sudoingX argued (76 likes, 12 replies, 5,785 views, 45 bookmarks) that a usable agent should start with almost nothing and let the operator add tools back one by one, because unused defaults waste context and clutter configuration. In the replies, people named the small set they actually want first: a persistent shell, web fetch, and repository-specific MCPs. The same dissatisfaction appeared under @AlexFinn said (198 likes, 34 replies, 11,478 views, 157 bookmarks), where a reply complained that users keep bouncing between Claude Code, Codex, and other top-tier platforms without a clean way to combine the best parts of each. The current workaround is to begin from a minimal base and rebuild only the pieces that prove necessary.
Trading-agent stacks are easy to assemble and hard to validate¶
Severity: Medium. @cyrilXBT shared (101 likes, 17 replies, 5,783 views, 85 bookmarks) a six-repo open stack for automated trading, but the most useful reply said Binance rate-limit errors consumed three days of debugging even for read-only sync. Another reply asked the direct question the rest of the thread avoided: whether anybody is actually making a profitable AI trader. @quantscience_ offered (34 likes, 4 replies, 3,124 views, 52 bookmarks) the more serious side of the space with a paper-backed framework and explicit analyst, risk, and portfolio-manager roles, which is itself a sign that ad hoc stacks are not enough. This looks worth building for, but buyers will expect stronger live-validation and risk-management evidence than most promotional threads currently provide.
3. What People Wish Existed¶
Persistent context that compounds instead of reloading every session¶
The clearest practical need was a memory layer that survives across tools, repos, and business systems. @GergelyOrosz described (110 likes, 25 replies, 10,297 views, 33 bookmarks) a rush of efforts to build exactly that for engineering teams, and a reply sharpened the requirement by distinguishing persistence from plain retrieval. @milesdeutscher shared (32 likes, 11 replies, 2,651 views, 41 bookmarks) an Obsidian-plus-Hermes setup that ingests client notes, SOPs, meeting logs, and business decisions into a "Business Brain" as a workaround today. This is a practical need with immediate workflow value. Opportunity: direct.
Portable, inspectable agent definitions and skills¶
Several of the strongest posts wanted the same property: an agent definition that can be read, edited, and moved around as files. @rauchg reduced (546 likes, 60 replies, 33,697 views, 240 bookmarks) an eve agent to instructions.md plus a skills folder; @hwchase17 pointed (49 likes, 6 replies, 3,965 views, 37 bookmarks) to Leve's filesystem-first LangGraph framework; and @sudoingX argued (76 likes, 12 replies, 5,785 views, 45 bookmarks) for adding tools back only when needed. @alex_prompter added (46 likes, 9 replies, 3,306 views, 89 bookmarks) evidence that behavior-only repos are now a real distribution channel. This is a practical need with multiple public implementations already shipping. Opportunity: direct.
Safer agent-native commerce and trading infrastructure¶
The emotional promise in the feed was autonomy; the practical request was safer execution. @quantscience_ used (34 likes, 4 replies, 3,124 views, 52 bookmarks) explicit analyst, researcher, risk, and portfolio-manager roles inside TradingAgents, which shows how much structure serious builders already need. @slash1sol described (98 likes, 28 replies, 5,222 views, 72 bookmarks) an agent marketplace with self-custody and direct settlement, while @Cryptic_Web3 highlighted (211 likes, 31 replies, 20,748 views) funding for agent-payment rails. The need is practical, but the solution space is already noisy and highly competitive. Opportunity: competitive.
4. Tools and Methods in Use¶
| Tool | Category | Sentiment | Strengths | Limitations |
|---|---|---|---|---|
| Claude Code | Coding agent | (+) | Central reference point for loop engineering, skills, and file-based agent definitions; also now appears on rabbit r1 surfaces | Users still need better harnesses, context layers, and token discipline around it |
| Codex | General-purpose agent | (+) | Browser and computer use let it finish full workflows such as GitHub push plus Vercel deploy | Users still juggle it alongside other top-tier agents rather than inside one unified stack |
| Hermes Agent | Agent runtime | (+/-) | Async subagents, iMessage support, automation blueprints, profile builder, atomic memory operations, and blank-slate setup | Fast-growing surface area creates demand for minimal installs and selective tool re-addition |
| eve | Agent framework | (+) | Extremely small markdown-and-skills definition, deployable on Vercel, with a typed hook for custom UIs | Evidence today is mostly early-product framing rather than long production case studies |
| Leve | Agent framework | (+) | Filesystem-first durable agents on LangGraph with approvals, schedules, subagents, and evals | Replies immediately asked about schema drift and concurrent directory mutations |
| Obsidian | Knowledge base | (+/-) | Gives agents a local vault for SOPs, meeting notes, and business context | Replies raised privacy discomfort about trusting AI with client notes and business records |
| GLM 5.2 | Model | (+/-) | Shows up in open-source capability claims and general-purpose app demos as a coding-capable open model | Much of today's supporting evidence was anecdotal or promotional rather than deeply benchmarked |
| TradingView Lightweight Charts | Charting library | (+) | Small, fast HTML5 financial charts for trading-agent frontends | Solves visualization only; execution and risk still require separate components |
| CCXT | Exchange API library | (+/-) | Unified API across 100+ exchanges for trading bots and data pipelines | Replies highlighted rate-limit and debugging pain in live usage |
| Tavily | Search API | (+) | Public site emphasizes security, privacy, and content-validation layers for live search | It is a supporting component, not a substitute for trading logic or execution controls |
The overall satisfaction spectrum was positive for modular, composable tools and mixed for big bundled runtimes. The common workaround pattern was to keep the agent definition in files, add tools only when needed, and use traces or explicit roles to keep long workflows understandable. Migration pressure is moving away from one giant default setup and toward portable behavior layers that can sit on top of Claude Code, Codex, Hermes, or a custom framework.
5. What People Are Building¶
| Project | Who built it | What it does | Problem it solves | Stack | Stage | Links |
|---|---|---|---|---|---|---|
| eve | @rauchg | Defines an agent as a tiny folder with instructions and skills, deployable on Vercel | Makes agent apps easier to inspect, share, and deploy | Markdown files, skills folder, Vercel | Beta | tweet |
| Leve | @jit_infinity | Compiles a directory of files into a durable LangGraph agent | Gives builders a filesystem-native way to add approvals, schedules, subagents, and evals | Python, LangGraph, file-system agent spec | Alpha | tweet, repo |
| Hermes Agent v0.17 | @IBuzovskyi | Expands a multi-channel runtime with iMessage, async subagents, automation blueprints, and a profile builder | Lets long-running agents span messaging, automation, and dashboard workflows without blocking the main session | Hermes runtime, dashboard, skills/MCP integrations, Grok OAuth | Shipped | tweet |
| TradingAgents | Tauric Research | Simulates a trading firm with analyst, researcher, trader, risk, and portfolio-manager agents | Adds explicit debate and risk structure to LLM trading workflows | Python, multi-agent LLM roles, benchmarked research framework | Beta | tweet, repo, paper |
| Tiny Place | TinyHumans AI | Creates a market surface where agents can discover bounties, negotiate fees, and settle in USDC | Gives agents a direct way to earn and transact instead of staying trapped inside one app | Solana, x402, Phantom, local agent stacks such as Hermes or OpenClaw | Shipped | tweet |
Leve and eve mattered together because they showed two versions of the same build pattern: keep the agent definition in plain files, then let the runtime compile or deploy around that source of truth. The difference is emphasis: eve made the accessibility argument, while Leve pushed harder on durability, approvals, schedules, and evals.
Hermes Agent's v0.17 release was the clearest evidence that runtime builders are racing to turn agent loops into full operating surfaces. iMessage support without a Mac relay, background subagents that return later, and blueprint-driven scheduling all point to agents being designed for longer-lived, multi-channel work rather than one terminal session.
TradingAgents was the strongest research-backed builder signal in the finance cluster. Its public repo and paper describe a team-of-agents structure with analysts, bullish and bearish researchers, risk management, and final portfolio approval, which is a much more explicit architecture than the simpler "connect six repos and go" recipes circulating alongside it.
Across these builds, the repeated pattern was explicit roles plus explicit artifacts: files, skills, schedules, traces, or risk layers. Even the commerce-oriented launch, Tiny Place, presented itself as a workflow surface with discovery, negotiation, and settlement steps, not just a tokenized slogan.
6. New and Notable¶
Skill-MAS treated orchestration knowledge as a text artifact instead of a weight update¶
@dair_ai shared (57 likes, 6 replies, 4,696 views, 56 bookmarks) Skill-MAS, and the public arXiv abstract for Skill-MAS says it tries to solve a recurring multi-agent problem: how to retain orchestration experience without fine-tuning the base frontier model. The paper's core move is to evolve a meta-skill in text using multi-trajectory rollout and selective reflection, which made it one of the day's clearest research signals around agent coordination rather than raw model scaling.
Public harness curriculum became part of the ecosystem¶
@rohit_tiwari surfaced (37 likes, 4 replies, 1,261 views, 36 bookmarks) Learn Harness Engineering as a free public course, while @alex_prompter highlighted (46 likes, 9 replies, 3,306 views, 89 bookmarks) a separate wave of skills-first and markdown-first repos. Together, those posts suggest that harness patterns are no longer only being discovered inside teams; they are now being packaged into teachable, copyable public artifacts.
7. Where the Opportunities Are¶
[+++] Durable context and verification layers for real engineering work — Evidence came from multiple sections: @GergelyOrosz described demand for engineering context layers, @Vtrivedy10 described the trace-and-eval burden inside harnesses, and Learn Harness Engineering packaged the same needs into a public curriculum. This is strong because the pain appears in practitioner workflow, market demand, and training material at the same time.
[++] Portable file-based agent packaging — The eve folder layout, Leve's filesystem-first framework, Hermes blank-slate setup, and skills-first repos all point toward a common need: agent definitions that are inspectable, editable, and transferable across runtimes. This is moderate because multiple approaches already exist, but the convergence itself suggests buyers care about the abstraction.
[+] Safer agent-native trading and payment operations — TradingAgents, cyrilXBT's open stack, Tiny Place, and aipaywithcrypto's funding round all show real energy around agents that can trade or pay. This is still emerging because the strongest unanswered questions in the discussion were about profitability, debugging burden, and live risk control rather than demand.
8. Takeaways¶
- Loop engineering is turning into a real operating discipline, not just a phrase. The day's strongest posts connected loops to harnesses, evals, context windows, and persistent memory rather than to prompts alone. (source)
- The preferred agent packaging format is getting simpler and more inspectable. Folder trees, markdown files, and skill directories showed up repeatedly across eve, Leve, and the skills-repo conversation. (source)
- General-purpose agents are increasingly judged by whether they can cross surfaces and finish operator work. Codex doing a full GitHub-plus-Vercel deploy, Hermes adding multi-channel runtime features, and rabbitOS carrying Claude Code onto r1 all fit that pattern. (source)
- Trading and commerce remain high-energy but mixed-signal. The serious evidence came from explicit frameworks and stack components, while replies kept pointing back to rate limits, debugging pain, and uncertainty about live profitability. (source)