Twitter AI Agent - 2026-06-09¶
1. What People Are Talking About¶
1.1 Harness engineering shifted from slogan to failure-closing infrastructure 🡕¶
The strongest June 9 theme was that harness engineering got more specific: the conversation moved from generic “loops, not prompts” talk to concrete verifier design, interruption handling, and self-improving harnesses. Four retained items supported this theme, and all four treated the harness as the main reliability surface rather than a thin wrapper around a model.
@akshay_pachaar argued (207 likes, 39 replies, 19,751 views, 263 bookmarks) that unattended loops do not remove debugging; they relocate it into runs no one watched, which means a loop is only useful if the checker can be trusted. The attached diagram made the point concrete by contrasting a manual trace-reading workflow with a self-closing loop that turns failures into root-cause analysis, proposed diffs, reruns, and locked regression tests.
@HowToAI_ framed (210 likes, 20 replies, 10,123 views, 211 bookmarks) the new Illinois, Meta, and Stanford survey as “Code as Agent Harness,” arguing that code is becoming the memory, environment, and boundary for agent systems rather than just their final output. The paper repository for Code as Agent Harness says the field is organizing around harness interfaces, harness mechanisms, and scaling the harness, which matches the day’s shift from prompt phrasing to executable and verifiable scaffolding.
@omarsar0 shared (76 likes, 10 replies, 4,033 views, 107 bookmarks) the new Self-Harness paper, which treats prompts, tools, and control flow as a learnable artifact instead of a hand-maintained scaffold. The preprint page shown in the image says the loop is weakness mining, harness proposal, and proposal validation, with held-out pass-rate gains reported across three base-model families.
@threepointone posted (49 likes, 4 replies, 2,905 views, 37 bookmarks) a “harness reliability engineering” state machine that mapped interruption outcomes such as RetryTurn, ParkForHuman, PreserveResult, RepairTranscript, and ReattachChild. That image mattered because it translated reliability work into explicit runtime transitions instead of vague advice.
Discussion insight: The useful pushback was that loops still fail silently unless something outside the loop notices and closes the failure. Replies on the Akshay thread stressed that debugging never disappeared; the system only became more complex about where debugging happens.
Comparison to prior day: June 8 made “loop engineering” mainstream shorthand. June 9 made the term more concrete by centering verifiers, self-improving harnesses, and interruption-state design.
1.2 Claude Fable 5 turned model quality into a routing and orchestration debate 🡕¶
The second major cluster was Anthropic’s Claude Fable 5 release, but the interesting part was not benchmark fandom by itself. Three retained items supported this theme, and together they turned the launch into an operating discussion about when to use frontier intelligence, when to delegate, and who should route the work.
@kimmonismus summarized (628 likes, 35 replies, 63,043 views, 132 bookmarks) Fable 5 as state-of-the-art across software engineering, knowledge work, vision, and scientific research, and the attached benchmark table showed the release-day headline numbers: 80.3% on SWE-Bench Pro, 29.3% on FrontierCode Diamond, and 88.0% on Terminal-Bench 2.1 for Claude Mythos 5 / Fable 5. Another attached chart compared cost and accuracy by effort level, making the trade-off itself part of the conversation.
@JJEnglert reviewed (161 likes, 17 replies, 28,936 views, 226 bookmarks) the model from a non-engineer operator perspective after spending roughly one billion tokens on real work. His thread said Fable found a serious bug missed by Claude 4.8 and GPT-5.5, but the same write-up also made the control problem explicit: he wants an auto-router that downshifts on simple work and only escalates to Fable for tasks that justify the cost.
@ClaudeDevs showed (107 likes, 1 reply, 33,449 views, 27 bookmarks) the release getting packaged immediately into a manager-worker pattern, with a coordinator agent on Claude Fable 5 delegating code review and test writing to smaller agents. The screenshot matters because it turned “use smaller models underneath” into concrete code rather than a vague strategy.
Discussion insight: The strongest nuance came from replies to both benchmark and review threads: some practitioners expected the biggest gains mainly on longer tasks and larger codebases, and even bullish users kept describing manual routing rather than trust in one model doing everything.
Comparison to prior day: June 8 already framed model choice as a routing problem. June 9 supplied the benchmark screenshots, operator economics, and coordinator code snippets that turned that thesis into release-day practice.
1.3 Workflow surfaces, skills, and delivery harnesses kept getting packaged into products 🡕¶
The third theme was packaging. Five retained items supported it, and they all pointed away from standalone chat shells and toward durable work surfaces: desktop runtimes, reusable workflow templates, portable skills, and delivery harnesses.
@karrisaarinen wrote (121 likes, 15 replies, 8,585 views, 71 bookmarks) that he wants an agent workflow tool that can describe work items, assign tasks, review code diffs, work multiplayer, collect customer context, define shared skills and MCP servers, and run from Slack. That post was the day’s clearest first-person product requirements list for an agent control plane.
@ollama announced (213 likes, 14 replies, 9,331 views, 99 bookmarks) that Hermes Desktop now runs through Ollama with a single command. The Hermes docs and Hermes Agent repo describe the same packaging move: desktop installers, 20-plus messaging surfaces, scheduled automations, MCP support, isolated subagents, and a built-in learning loop.
@tonbistudio shared (39 likes, 2 replies, 3,006 views, 49 bookmarks) an open-source Hermes multi-agent workflow template whose README lays out a fixed shape of intake, dedup, scoring, parallel research, routing, one human gate, and fulfillment. @Coo_Lxing introduced (16 likes, 853 views, 15 bookmarks) Loom as a delivery harness for planning, verification, repair, preview, and handoff that survives interruption and compaction.
@twostraws shared (98 likes, 4 replies, 3,095 views, 20 bookmarks) that Xcode featured one of his agent skills, and the Swift Testing Agent Skill repo shows install paths across Xcode, Claude Code, Codex, Cursor, and Gemini. @LangChain added (15 likes, 2 replies, 1,780 views, 12 bookmarks) another layer with interpreter skills, where a skill ships with a TypeScript module the runtime can import and execute.
Discussion insight: The shared request was not “give me one smarter chat box.” It was “give me a place where plans, skills, tools, context, verification, and handoff can persist.”
Comparison to prior day: June 8 already had skills in Slack, Xcode, and deployment stacks. June 9 broadened that packaging into desktop runtimes, executable skills, reusable triage templates, and delivery harnesses.
2. What Frustrates People¶
Silent failure in unattended loops¶
Severity: High. @akshay_pachaar argued (207 likes, 39 replies, 19,751 views, 263 bookmarks) that unattended loops only help if the checker can be trusted, because otherwise the human is back to reading traces by hand. @threepointone posted (49 likes, 4 replies, 2,905 views, 37 bookmarks) a state machine full of interruption branches, which showed how much failure handling sits outside the happy path. The Loom README names the same breakdowns directly: partial completion, goal drift, self-check bias, token waste, and handoff gaps. People cope by building repair loops, human parking states, and regression locks. This looks worth building for because the feed treated silent failure as infrastructure debt, not user error.
Long-running agent work still lacks a durable control plane¶
Severity: High. @karrisaarinen wrote (121 likes, 15 replies, 8,585 views, 71 bookmarks) a requirements list for a workflow tool spanning plans, task assignment, diff review, multiplayer collaboration, shared skills, shared MCP servers, and Slack-native operation. The Hermes multi-agent workflow template and Loom both address parts of that gap with one-human-gate routing, persisted task state, and delivery evidence, but both present themselves as structures to adapt rather than turnkey control planes. Current workarounds are templates, Kanban boards, and custom harnesses. The need looks strong because the complaints were about continuity, inspectability, and shared context rather than raw model intelligence.
Manual routing is still the tax on frontier-model performance¶
Severity: High. @JJEnglert reviewed (161 likes, 17 replies, 28,936 views, 226 bookmarks) Claude Fable 5 as worth the spend on hard work, but explicitly asked for an auto-router that saves frontier reasoning for tasks that earn it. @ClaudeDevs showed (107 likes, 1 reply, 33,449 views, 27 bookmarks) one workaround by delegating subtasks to smaller agents, while replies to @kimmonismus summarizing (628 likes, 35 replies, 63,043 views, 132 bookmarks) Fable 5 said the gains may be most obvious on larger codebases and longer runs. The workaround is hand-built relays between frontier planners and cheaper executors. That makes this worth building for because the value is visible, but the routing logic still lives in operator judgment.
MCP and enterprise tool exposure still look too broad by default¶
Severity: Medium. @saketsaurabh launched (98 likes, 10 replies, 3,357 views, 14 bookmarks) MCP Studio with the argument that one-MCP-server-per-app creates too many tools, too many tokens, and too many governance problems. His thread said real enterprise tasks cut across Salesforce, Snowflake, NetSuite, ServiceNow, Workday, and internal systems, so mirroring every application into one tool pile strains production use. The current workaround is manual tool curation or task-specific setup. This looks worth building for because the complaint is about scope control and governance, not feature completeness.
3. What People Wish Existed¶
A shared workflow layer for plans, diffs, context, and team collaboration¶
This was the clearest direct need on June 9. @karrisaarinen wrote (121 likes, 15 replies, 8,585 views, 71 bookmarks) that he wants one tool to describe work items, assign tasks to agents, review diffs, work multiplayer, collect customer context, define shared skills and MCP servers, and run from Slack. The Hermes multi-agent workflow and Loom repos show partial answers today, but both still assume users will adapt the structure to their own process. Opportunity: direct.
Automatic routing between frontier planners and cheaper workers¶
This need was urgent and operational. @JJEnglert reviewed (161 likes, 17 replies, 28,936 views, 226 bookmarks) Fable 5 positively while still asking for an auto-router that downshifts on simple work and escalates only when the task earns frontier spend. @ClaudeDevs showed (107 likes, 1 reply, 33,449 views, 27 bookmarks) the manual version already happening through coordinator-plus-subagent orchestration, and @kimmonismus summarized (628 likes, 35 replies, 63,043 views, 132 bookmarks) benchmark gains that make the routing decision economically meaningful. Opportunity: direct.
Delivery harnesses that survive interruption, compaction, and handoff¶
This need was practical rather than aspirational. @Coo_Lxing introduced (16 likes, 853 views, 15 bookmarks) Loom as a delivery harness for planning, verification, repair, preview, and handoff, and the repo says the goal is to preserve project context, test results, repair notes, preview evidence, and delivery reports so the next session or agent can continue without restarting. @akshay_pachaar argued (207 likes, 39 replies, 19,751 views, 263 bookmarks) that failure-closing logic has to be built explicitly, while @threepointone posted (49 likes, 4 replies, 2,905 views, 37 bookmarks) the interruption-state view of the same problem. Opportunity: direct.
Portable skills that carry executable behavior, not just instructions¶
This need looked competitive but real. @twostraws shared (98 likes, 4 replies, 3,095 views, 20 bookmarks) a skill that can be installed across Xcode and other coding agents, and the Swift Testing Agent Skill repo documents those multi-surface install paths directly. @LangChain added (15 likes, 2 replies, 1,780 views, 12 bookmarks) interpreter skills, where SKILL.md ships alongside a TypeScript module the interpreter can import and run. The request underneath both examples is clear: reusable skill packages should be executable, portable, and versionable. Opportunity: competitive.
4. Tools and Methods in Use¶
| Tool | Category | Sentiment | Strengths | Limitations |
|---|---|---|---|---|
| Harness engineering / self-closing loops | Method | (+) | Turns traces into root-cause analysis, reruns, and regression locks | Still depends on trusted checkers and explicit failure handling |
| Code as Agent Harness | Research / design pattern | (+) | Treats code as executable memory, environment, and verification substrate | Research framing rather than a turnkey product |
| Claude Fable 5 | Frontier LLM | (+/-) | Strong benchmark showing on coding and long-horizon work; catches issues other models miss | Token spend stays high and routing is still manual |
| Claude Managed Agents pattern | Multi-agent orchestration | (+) | Lets a frontier coordinator delegate review and test-writing work to smaller agents | Requires operator judgment about agent split and task routing |
| Hermes Desktop + Ollama | Desktop runtime | (+) | Local-friendly GUI, subagents, skills, messaging integrations, MCP support | Product complexity is higher than a single-terminal coding workflow |
| Hermes multi-agent workflow | Orchestration template | (+) | Gives teams a reusable intake-to-research-to-human-gate pipeline | Template only; still needs domain adaptation and ops setup |
| Loom | Delivery harness | (+) | Persists context, verification evidence, repair notes, preview checks, and handoff state | Early project and extra process overhead for small one-shot edits |
| LangChain interpreter skills | Skill packaging | (+) | Adds executable TypeScript modules to skills for reusable behavior | Requires interpreter support and deliberate exposure controls |
| Swift Testing Agent Skill | Agent skill | (+) | Packages high-signal Swift Testing guidance for Xcode and multiple coding agents | Narrow domain; value depends on teams already using skills |
| MCP Studio | Tool gateway / MCP | (+/-) | Narrows tool scope around tasks instead of mirroring every app into one server | Still has to prove governance and usability across messy enterprise stacks |
| Cantrip | Agent runtime | (+/-) | Explicit boundaries, replayable looms, child composition, and ACP support | New runtime concepts may be powerful but add adoption friction |
The overall satisfaction spectrum was pragmatic. People were enthusiastic about methods and products that make agents more inspectable, resumable, and packageable rather than simply “more autonomous.” The main migration pattern was to keep frontier models on planning or review, push bounded work into smaller workers or reusable workflows, and move long-horizon continuity into harnesses, skills, or desktop runtimes (@JJEnglert reviewed (161 likes, 17 replies, 28,936 views, 226 bookmarks); @ClaudeDevs showed (107 likes, 1 reply, 33,449 views, 27 bookmarks); Hermes docs; Loom; Hermes multi-agent workflow). The competitive dynamic is shifting toward the best control layer around the model.
5. What People Are Building¶
| Project | Who built it | What it does | Problem it solves | Stack | Stage | Links |
|---|---|---|---|---|---|---|
| Hermes multi-agent workflow | @tonbistudio | Reusable autonomous triage pipeline with intake, dedup, scoring, parallel research, routing, one human gate, and fulfillment | Gives teams a concrete orchestration skeleton instead of inventing a workflow from scratch | Python, YAML triage.yaml, Hermes Kanban |
Shipped | repo, tweet |
| Loom | @Coo_Lxing | Open-source delivery harness for coding agents with planning, verification, repair, preview, and handoff state | Keeps long-running delivery work aligned across interruptions, compaction, and handoffs | Agent-neutral CLI, dynamic workflows, persisted .loom/ state |
Alpha | repo, tweet |
| Hermes Desktop | Nous Research | Desktop surface for Hermes Agent with local or cloud models, subagents, skills, and messaging integrations | Packages a self-improving agent runtime into a GUI and local-friendly install flow | Desktop app, Hermes Agent runtime, Ollama, MCP, subagents | Shipped | docs, repo, tweet |
| Swift Testing Agent Skill | @twostraws | Portable skill for writing and improving Swift tests across Xcode and coding agents | Packages domain expertise into an installable skill instead of repeating prompt instructions | Swift skill package, Xcode plug-ins, Claude Code/Codex/Cursor/Gemini installs | Shipped | repo, tweet |
| Cantrip | deepfates | Elixir agent runtime organized around circles, media, gates, wards, looms, and child composition | Gives agent builders explicit boundaries, replay, persistent state, and ACP access | Elixir, OTP processes, JSONL/Mnesia loom storage, ACP | Alpha | repo, tweet |
| MCP Studio | @saketsaurabh | Task-oriented MCP layer that aims to bring the right tools across systems together for one job | Reduces tool sprawl, token overhead, and governance issues from one-server-per-app MCP setups | MCP, enterprise SaaS connectors, task-specific tool routing | Beta | tweet |
The common build pattern was not “one agent to rule them all.” It was “externalize the coordination layer.” Hermes multi-agent workflow and Loom both turn agent work into a staged delivery process with explicit gates, while Hermes Desktop and Cantrip package memory, delegation, and runtime boundaries into reusable surfaces.
Swift Testing Agent Skill and LangChain’s interpreter-skills direction point to another repeated pattern: expertise is getting shipped as portable modules. That matters because the feed kept rewarding projects that make behavior installable, reviewable, and reusable rather than re-explained on every run.
MCP Studio fits the same trend from the enterprise side. Instead of exposing every tool from every system, it argues for composing the right subset of tools around each task, which is another form of workflow packaging.
6. New and Notable¶
Code-as-harness research became the day’s clearest conceptual anchor¶
@HowToAI_ framed (210 likes, 20 replies, 10,123 views, 211 bookmarks) the Illinois, Meta, and Stanford survey as a shift from natural-language reasoning to code as memory, environment, and boundary. The public Code as Agent Harness repository says the field is now being organized around harness interfaces, harness mechanisms, and scaling the harness. This mattered because it gave a name and a structure to the broader harness conversation already building across the feed.
Release-day Fable benchmark images sharpened the coding-agent race¶
@kimmonismus summarized (628 likes, 35 replies, 63,043 views, 132 bookmarks) Claude Fable 5 with benchmark screenshots that put concrete numbers behind the model conversation: 80.3% on SWE-Bench Pro, 29.3% on FrontierCode Diamond, and 88.0% on Terminal-Bench 2.1 in the attached table. This mattered because the day’s model debate stopped being abstract almost immediately and turned into workload, effort-setting, and routing decisions.
Skills gained executable code attachments, not just markdown instructions¶
@LangChain added (15 likes, 2 replies, 1,780 views, 12 bookmarks) interpreter skills as a new extension to agent skills. The public interpreter skills write-up explains the shift: a skill can now ship with a TypeScript module that the interpreter imports and runs, including subagent and tool calls under explicit runtime exposure controls. This matters because it turns skills into both an instruction surface and an API surface.
7. Where the Opportunities Are¶
[+++] Agent delivery control planes — The strongest evidence came from multiple sections at once: @karrisaarinen wrote (121 likes, 15 replies, 8,585 views, 71 bookmarks) a direct spec for a workflow tool; @akshay_pachaar argued (207 likes, 39 replies, 19,751 views, 263 bookmarks) that loops need trusted failure-closing logic; Loom is explicitly building for interruption, verification, repair, and handoff; and Hermes multi-agent workflow packages a one-human-gate pipeline. The gap is strong because the pieces exist, but the shared control plane still looks fragmented.
[++] Frontier-model routing and delegated worker orchestration — @JJEnglert reviewed (161 likes, 17 replies, 28,936 views, 226 bookmarks) Fable 5 as valuable but still too manual to route, @ClaudeDevs showed (107 likes, 1 reply, 33,449 views, 27 bookmarks) a coordinator-plus-subagent pattern, and @kimmonismus summarized (628 likes, 35 replies, 63,043 views, 132 bookmarks) benchmark gains large enough to justify routing discipline. This is moderate because the economics are obvious, but the policy still lives mostly in human heads.
[++] Portable executable skills and workflow artifacts — @twostraws shared (98 likes, 4 replies, 3,095 views, 20 bookmarks) a skill that installs across several agent surfaces, @LangChain added (15 likes, 2 replies, 1,780 views, 12 bookmarks) executable interpreter skills, and @tonbistudio shared (39 likes, 2 replies, 3,006 views, 49 bookmarks) a reusable triage template. This is moderate because the packaging pattern is clear, but standards and distribution are still early.
[+] Task-scoped MCP and enterprise tool-governance layers — @saketsaurabh launched (98 likes, 10 replies, 3,357 views, 14 bookmarks) MCP Studio by arguing that one-MCP-server-per-app creates too many tools, tokens, and governance issues. This is emerging because the pain is concrete, but the category still looks new and vendor-specific.
8. Takeaways¶
- Harness talk got more operational today. @akshay_pachaar argued (207 likes, 39 replies, 19,751 views, 263 bookmarks) that loops only help if the checker is trustworthy, while @threepointone posted (49 likes, 4 replies, 2,905 views, 37 bookmarks) an interruption-handling state machine. (source)
- The research framing behind the shift is now clearer. @HowToAI_ framed (210 likes, 20 replies, 10,123 views, 211 bookmarks) code as the harness, and the public Code as Agent Harness repo organizes that idea into interfaces, mechanisms, and scaling concerns. (source)
- Fable 5 made routing discipline more important, not less. @kimmonismus summarized (628 likes, 35 replies, 63,043 views, 132 bookmarks) benchmark gains, but @JJEnglert reviewed (161 likes, 17 replies, 28,936 views, 226 bookmarks) the same release while still asking for an auto-router. (source)
- The missing product is a durable agent workflow layer. @karrisaarinen wrote (121 likes, 15 replies, 8,585 views, 71 bookmarks) the most explicit requirement list of the day, and both Loom and Hermes multi-agent workflow are trying to fill parts of that gap. (source)
- Skills are turning into executable, portable assets. @twostraws shared (98 likes, 4 replies, 3,095 views, 20 bookmarks) a cross-surface skill install, while @LangChain added (15 likes, 2 replies, 1,780 views, 12 bookmarks) interpreter skills that bundle runnable code with instructions. (source)