BlogDeep-dives into AI infrastructure, prompt engineering, and the tools we build.
Hermes Agent's memory/skills guidance dropped from 537 to 255 tokens per call (-282) by removing instructions the memory tool schema already teaches.
Hermes's clarify tool now asks 2-5 questions per call instead of one, cutting agent-human round-trips and the 16-27 option combinatorial blowup.
Hermes plugins render live inline UI in the chat transcript. Agents emit ::directive lines, plugins claim the name, output paints in a sandboxed frame.
continuity=true injects a cron job's own 8K-truncated last output into each run, so recurring jobs dedupe against what they already reported.
Agent Plugins 1.0: a portable format for agent skills and MCP servers, backed by AWS, Cursor, Microsoft, OpenAI, and Vercel. Hermes shipped support same day.
Fail-closed shell hooks, AGENTS.md directory-chain merging, slash-menu fuzzy scoring, and --resume latest arrive in Hermes. GitHub-sourced (X API paywalled).
Cline injected Meta's Muse Spark 1.2 harness principles into their agent prompt. Same model, same task: 19.7M tokens to 7.2M, 49 minutes to 24.
/learn, /refine, /heartbeat, and quality gates land in a single 24-hour window alongside STT hardening. All sourced from GitHub - X API paywalled.
Hermes Agent merged three PRs today: quality gates verify /goal output, /heartbeat re-engages idle sessions, and /refine reviews memory on demand.
Three Prime-Agent features landed: on-demand memory review, recurring idle session prompts, and deterministic goal verification for Hermes Agent.
zx_joy_ shares 184 Hermes skill files with Claude Code via symlinks. One rule in CLAUDE.md. Bidirectional updates. Zero MCP servers, zero API.
v0.20.0 adds streaming conversational voice with barge-in, on-device wake words, signed outbound webhooks, grounded citations, and A2A v1.0 agent protocol.
Hermes v0.20.0: streaming voice with barge-in, wake words, A2A v1.0, self-healing tools, desktop plugin SDK. 3,650 commits, 1,400 PRs, 650+ contributors.
Hermes Agent system prompts burn 600+ tokens by default. Community optimization cuts that to 120. The skills index alone costs 1.5K-2.5K per turn.
Kanban cards can override model and reasoning effort. The model picker switches from a blocking 7.3s payload build to 0.06s stale-while-revalidate.
The 12K-token tool schema was re-sent uncached on every Anthropic call, eating 70% of input spend. A new 4-breakpoint cache plan stops the waste.
PR #76032 re-plans Anthropic's cache budget so the 12K-token tool schema becomes a cached prefix, cutting ~70% of input spend on tool-heavy agents.
n8n handles the deterministic backbone; Hermes handles reasoning. Four consolidated webhooks replace 15+ tools, cutting tool-description tokens by 73% per turn.
Per-turn micro-compaction lands in Hermes, idle CPU fix saves 20% background burn, FLUX 3 video gen ships, 5 desktop UX PRs land.
Two model-routing plugins shipped for Hermes Agent within 48 hours of Liquid AI encoder release. Classifies by sensitivity, role, and difficulty.
Vercel AI Gateway and Vercel Sandbox return in a 4,259-line PR. iMessage-style emoji reactions land on desktop. Comfy Cloud joins MCP catalog.
Hermes Crystal compactor reduced 4.17M source tokens to 262K in a 4-hour session — 93.7% reduction with zero continuity breaks, 11ms compactions.
A performance sprint on cold-start, voice full-duplex interrupt, server-owned pins, and a 6,316-line CLI decomposition -- sourced from GitHub while X paywalled
L0-L3 layered memory with per-block 6,000-char budgets prevents context explosion. Combined with a token-disciplined system prompt for max recall per token.
Three perf PRs target token accounting, HTTP connection reuse, and SQL query optimization. Combined 2,046 additions across 17 files, salvaging community work.
New Hermes plugin rewrites prompts per model family before they hit the LLM, averaging +55 quality points and cutting tokens 20-80% per rewrite.
Hermes ports Claude Code's session-wide runaway-loop caps, Desktop hits 60fps render, and a public subagent lifecycle API lands for plugin authors.
Hermes v0.19 picks up from the exact turn that crashed, checks what already happened, and finishes the job with no duplication.
4 desktop perf PRs merged July 26: streaming cost decoupled from transcript length, deferred timeline, faster model picker, and multitab streaming.
Iron-proxy mints opaque stand-in tokens so sandboxed agents never see real credentials. Compromised sandbox = stolen keys are useless.
Relay Phases 1+2 land media parity (+1,344 LOC), desktop connects to remote gateways, Arabic RTL ships, and session model locking salvages a 3-month-old PR.
Monolithic agent calls spend 80% of tokens on steps a bash script could handle. Graph routing with a classifier node yields 5-10x cost savings.
Hermes Agent merged 51 feature/perf PRs in 48 hours. Org-skills namespace with token-gated discovery pairs with two new ContextEngine plugin lifecycle verbs.
Most Hermes cron ticks need zero reasoning. The SILENT pattern skips LLM calls on idle cycles. Combined with tiered routing, users report ~60% cost reduction.
A compression overhaul: absolute token thresholds, idle-triggered compaction, structured logging. +1,162 lines, 21 files across three PRs by teknium1.
Hermes Agent has 11 auxiliary model slots for offloading side tasks. Override them with cheap flash models to cut non-reasoning costs by 90%+.
Hermes v0.19.0 with 2,245 commits from 450+ contributors. New: cross-surface theme SDK, command secret sources, and per-model compression thresholds.
v0.19.0 cuts cold-start latency from 4.3s to 0.9s, adds durable response delivery via state.db, and streams subagent transcripts live.
OutThisLife shipped 7 desktop perf PRs on July 20: virtualized diffs, targeted file-tree revalidation, rAF-coalesced resizes. Teknium1 bundled office skills.
A CLI audits Hermes Agent token burn, applies presets to disable unused tools and skills, cutting fixed overhead 32% and per-turn growth 96%.
Hermes Agent main absorbed 52 feature/perf PRs in 48h. @OutThisLife shipped 8 desktop rendering PRs targeting re-render and serialize bottlenecks.
Pinning the session-context render and moving per-turn notes to the user message raised Hermes Agent's prompt cache hit rate from 0.1% to 98.7%.
Byte-stable system prompts and an api_content sidecar eliminate the first-call cache miss, cutting response time from 27.9s to 2.4-5.8s on 156k-token contexts.
Terminal billing lands, desktop UX hardens, and Codex gains real-time observability -- all sourced from GitHub while X remains paywalled.
Three PRs added per-model, per-aux-task, and per-MoA-slot reasoning effort controls to Hermes. Side tasks at low, main model stays at full power.
A 4-phase plugin expansion plan lands as a community tracking issue. Kanban gets attachments, desktop rebindable keys, Unreal Engine joins MCP.
Blender MCP joins the Hermes catalog. `hermes mcp install blender` ships a curated 4-tool surface. Three PRs salvaged 620 lines of community reference material.
NouResearch hackathon winners: Custodian (spending kernel), Mom-n-Pop Skills (small business ops), CashFromChaos (photo-to-sale). 292 submissions, 3 winners.
teknium1 files a 4-phase plan for a production-grade Hermes plugin system. API versioning, inter-plugin event bus, pluggable approval, and a community index.
4x RTX 3090 Hermes Agent rig doubled client net profit. Plus: Grok 4.5 content machine, local QVAC deployment, restaurant booking API.
A community skill collection brings multi-agent Kanban orchestration to Hermes Agent with 36 composable skills and 7 role-specific agents.
6 feature PRs in 48h: durable background delegations survive restarts, Codex usage reset lands, DeepInfra provider ships, max/ultra reasoning effort levels.
Hermes gateway now reports 8 subsystem checks through /health/detailed. SQLite read-only probe, disk pressure, platform state, queue depth - no SSH required.
Six feature PRs in 48h: desktop approval mode, session import flow, max/ultra reasoning levels, and CLI workspace binding.
Hermes Agent uses dual-layer compression and a 4-breakpoint caching strategy to cap context length, cutting tokens by up to 75% on Anthropic.
July 12 roundup: max/ultra reasoning effort across all surfaces, CLI workspace filtering, gateway readiness checks, and Fireworks AI as a bundled provider.
Agent Install Monitor records every package install, Docker pull, and git clone executed through Hermes Agent into a local SQLite database for later inspection.
Hermes Agent on rabbitOS. Desktop cloud discovery. Community: rabbit r1 voice routing, Grok 4.5 cost analysis, install-monitor plugin.
GPT-5.6 lands in Hermes. HY3 free via Nous Portal. Clacky adds Hermes background agents.
Three PRs merged July 7 deliver 6 export formats, unified filters, and direct HF dataset upload. 3,901 lines of new session portability.
Teknium confirms day-zero Hermes Agent support for GPT-5.6. Plus: Shilong Liu's agent taxonomy, IBuzovskyi's 10-command breakdown.
Post-compression token estimates excluded the tools schema, causing a 20-30K blind spot. The fix was a one-line change.
Buildathon hits SF this week, CuTeDSL skill from Teknium's team, Unbroker skill hits 1,200 bookmarks, and a 4-machine production case study.
Progressive tool disclosure hides deferred MCP schemas behind three bridge tools. BM25 retrieval cuts 22K tokens per turn, Opus 4 accuracy jumps from 49% to 74%.
A Chinese founder runs his business on 4 Mac Mini Studios with Hermes Agent, no cloud fees. Plus: AI clone with Venice API, SecretSource commits, Arabic guide.
PR #50138 ships 8 files / +4,372 LOC for a record, generate, replay pipeline using CGEventTap. Self-learning mode scores 1.46 vs 3.01.
$750K Hermes Buildathon (10 cities, 5 countries) hit 202L/90B/83K imp. v0.14 cost stories and OpenClaw switchouts lead the day.
Eight Kanban SQLite corruption issues fixed in one week. WAL sidecar unlink races, SQLITE_BUSY aborts, torn writes. Six commits harden the surface.
Five posts in 24h shifted Hermes talk from hype to production: a studio on four machines, local hardware, gateway search, hermes-coder profile, UNBROKER skill.
Six builders shipped fresh Hermes v0.18 implementations in 24h: Rust CDP plugin, MCP testimonial, /bioinformatics command, memory how-to, and setup guide.
v0.18 Judgement: 2,333 likes, 219 RTs, 626 bookmarks, 639K impressions in 24h. Token-cost walkthroughs, offline brains, Nous tool gateway fill the rest.
PR #56254 adds cron_continuable_surface so Hermes cron delivers flat into Slack channels. Generic capability flag pattern, 13 files, +1115/-1, 280 tests pass.
Wandermist's 10x-faster article: 121 likes, 352 bookmarks, 141K impressions in 24h. JPIE's browser extension hit 300+ stars in 5 days, 4 releases shipped.
imbabybrooklyn's living map turns every Hermes agent's accumulated memory and skills into a shareable, animated graph. 85 likes, 39 bookmarks in hours.
Hermes Hackathon final day. @NousResearch 328L deadline, @wandermist 10x-speed article, 3 day-one agent submissions (AgentInvoice, Clawmes, Talaria).
Synapse: Graphiti + FalkorDB Hermes memory plugin with salience-based decay and valid_from/valid_to edges. Low-relevance facts expire, key ones persist.
XONA, MIND Protocol, PIO insurance, Lumina: day-one entries to the Hermes hackathon built on NVIDIA NIM, Stripe, and the agent loop.
Hermes Agent ships MoA presets as a virtual model provider with configurable references, aggregator budgets, and provider routing. 30+ commits in 72h.
June 27: MoA 2.0 crossed 4,500 likes in 17h, @Teknium's technical breakdown hit 2,000 likes, @mr_r0b0t's /learn trick landed 22K impressions.
Three independent developers shipped Chrome browser companions for Hermes Agent between April 20 and June 25. All connect to the local API on port 8642.
June 26: DeepSeek V4 wired into Hermes Agent (21 likes), Squadic Agentic Money Operator for the hackathon, GLM-5.2 free on Cloudflare Workers AI.
Hermes swapped cron prompt_cache_key from per-fire session_id to sha256 of the static prefix. Sub-TTL cron fires now reuse the warm prefix.
June 25: /learn explainers sweep X, Boxmining's walkthrough clears 4,288 impressions, PR #52273 caps desktop renderer at 20K chars to stop freeze.
Hermes Agent's /learn turns directories, URLs, conversations, and pasted text into a verified SKILL.md. 11 files, +404 LOC, zero new ingestion engine.
June 24 roundup: Telegram receipt-split bot, Slack as Hermes frontend, /learn for directories, real estate agent saves $18k/year.
auxiliary.background_review routes the post-turn review to a cheaper model. 39x cheaper, drops 1 skill and 1 memory on a real 102K session. Prompt-cache aware.
Two PRs in 24h: llm.oneshot gateway RPC bypasses session history (+347 LOC) and a desktop preflight checks cua-driver on macOS, Windows, and Linux (+616 LOC).
Hermes v0.17 ships a gRPC-native iMessage channel via Photon. No Mac relay, no public URL. Two commands to first text; persistent stream replaces webhooks.
June 22 roundup: Blank Slate mode hit 2,902 likes, manavv announced a Hermes Unbundled three-part series, community debated context compression at 32K tokens.
delegate_task(background=true) dispatches async subagents. Main agent keeps working. 1,268 lines added.
Sunday's Hermes Agent discussions: Desktop hack videos, OpenClaw-to-Hermes GPU farm migration, Obsidian integration, and a new open-source email tool for your agent.
Hermes v0.17.0 shipped 800 PRs, iMessage via Photon, and blank-slate profiles. Community built fleets, remote setups, and Grok provider hacks within days.
A Hermes Agent plugin rewrites prompts per-model, adding 55 quality points and cutting tokens 20–80% without user effort. SQLite analytics and slash commands for insights.
Hermes Agent is becoming a platform for agent collaboration, automation, and custom workflows. Top posts from the last 24h: A2A plugins, v0.17.0 release (browser control, MCP, Openclaw), and tutorials on building agents from scratch.
cobi_bean's framework: pick memory layers by forgetting cost, not plugin availability. Built-in memory, Honcho, and skills each handle a different failure.
The Hermes Agent Curator no longer auto-consolidates. The pipeline cost up to $16/week on frontier models. Capafy launched a Hermes skills marketplace.
Hermes Agent adds three optional payments skills via Stripe: virtual cards, HTTP 402 per-request API payments, and SaaS provisioning with spend controls.
A Chinese Hermes Agent tutorial hit 70 likes and 131 bookmarks in 3 days. Community also tackling llm-wiki scaling and corporate Mattermost deployments.
June 16 roundup: OpenGradient Veil keeps agent prompts private, AgentBus hits 100 holders, Box positions Hermes as the company brain, Stripe skills reaction continues.
Hermes Agent bundles an LLM Wiki skill that implements Karpathy's interlinked markdown knowledge base. Ingest, query, lint, and Obsidian sync built in.
Hermes Agent community roundup - OpenClaw production comparisons, Bangalore builder event, DeepSeek-v4-pro compatibility, Nora agent ops platform.
PM roundup: dual agent setups (Hermes VPS + OpenClaw laptop), a practitioner switching from OpenClaw to Hermes, Desktop one-command launch tutorial, and an iOS companion app.
14 parameterized blueprints render natively on every Hermes surface - dashboard forms, slash commands, or conversational fill. No cron syntax required.
Teknium's Manim video generation demo hit 108 likes and 3,100+ impressions. HYRE Gateway shipped 11-model routing. More users migrating from ChatGPT and Claude.
Hyre Gateway published an integration guide for Hermes behind 11-model routing. Cloudzy listed Hermes on its VPS marketplace. Two community guides dropped.
Hermes Agent's new Automation Blueprints turn cron jobs into conversational workflows with three trigger types: Schedule, GitHub Event, and API Call.
Teknium shipped Telegram Rich Messages for Hermes Agent, enabling markdown output, inline images, and PDF delivery directly in Telegram chat.
14 parameterized automation templates with typed slots render natively on every Hermes surface - dashboard forms, CLI slash commands, and agent conversations.
Token Lens plugin, Hermes Desktop app, CREAO AI vs Hermes comparison, Codex self-evolution analysis, and a mobile access PR from the community.
Headroom compresses agent context by 87% with no accuracy loss. 60-95% fewer tokens. MCP server gives Hermes Agent single-command integration.
A complete Hermes Agent guide goes viral in Chinese. Agent profiles, Mystery Shopper skills, and a 4-Mac-Mini local agent farm round out the day's top posts.
Headroom compresses tool outputs, logs, files, and RAG chunks before reaching the LLM. 60-95% fewer tokens, same answers. MCP server for Hermes Agent.
VPS remote desktop workflows drive the week. Five commits land in 24h. Multi-backend support signals an ecosystem maturing past provider lock-in.
Hermes ecosystem roundup: harness-as-moat thesis, ElevenLabs voice integration, model-task-routing skills, retail agent use cases, and in-person meetups.
Environment-gated skills, built-in auto-pruning after 90 days, and blank-slate installs shrink Hermes v0.16's default skill surface.
gentle-ai v1.37.0 ships full Hermes support. A community PR fixes Langfuse token tracking. Ollama arrives. Desktop GUI gains tool backend management.
Headroom compresses tool outputs, logs, files, and RAG chunks before they reach the LLM. 60-95% fewer tokens with accuracy preserved on GSM8K and TruthfulQA.
Three models on the same data gave WAIC scores of -2103, -1742, and -1707. The difference: how they handled 30 groups within 600 observations.
Community builds: a skill stress-test plugin, a Twilio voice agent, and cheap-model multi-profile orchestration on Hermes.
Hermes reads Obsidian vaults as live context, creating a second brain that talks back. 2,849 bookmarks on X and $35 hardware builds prove the pattern works.
Weekly X roundup: $4,400/mo from self-hosted Hermes, full agent courses emerge, new Labyrinth observability tool, and OpenClaw migration accelerates.
Hermes Agent's Obsidian plugin turns a passive second brain into an active reasoning partner. Notes, backlinks, and wikilinks become live agent context.
Monetization article hits 12.5K impressions, memory layer narrative dominates, mid-turn steering ships as desktop feature, skill PR slashes context 93%.
Multi-profile gateway sockets, RunbookHermes AIOps layer, KongBrain memory for OpenClaw, and Hermes + ElevenLabs phone integration.
Hermes Agent defers MCP tools behind three bridge tools at 10% context threshold. BM25 retrieval handles 80+ tool catalogs with ~300 token fixed overhead.
Warden Protocol adds x402 payments to Hermes for P2P AI inference. Hermes Desktop hits 10K stars. One user logs 663M tokens in 30 days.
hermesd is a read-only TUI dashboard that monitors Hermes Agent - sessions, tokens, cron, logs - from a single terminal window without writing to agent state.
Hermes Desktop public preview for Mac/Windows/Linux. Microsoft ships Scout on OpenClaw. Community builds agent harnesses and skill marketplaces.
Tool Search hides MCP and plugin tools behind BM25 bridges in Hermes Agent. 226-tool catalog drops from 53,994 to 2,289 tokens - a 95.8% reduction.
Ara launched an IDE powered by Hermes Agent, Desktop v1.0 arrived after 1.8k stars and a month of dogfooding, and the community wants more tutorials.
ComfyUI's Hermes Agent design workflow hits 3.1K impressions. Julian Goldie SEO's Hermes+Obsidian memory stack gains 1.9K impressions. Elkim ships Fluxer Discord-like plugin. Memory architecture dominates conversation.
Tool Search cuts context from 41% to 3%. Skills self-patch on mismatch. Agent capabilities are becoming dynamic and self-maintaining.
Same concept, 13 diagram-as-code tools, one API. Mermaid, D2, PlantUML, Graphviz, Excalidraw compared side by side through Kroki.
RTX Spark partnership hits 195 likes. Memory architecture deep dives detail Hermes's layered design: two capped files, frozen prompt, compression flush.
Hermes Agent runs fully local on Qwen 3.6 27B. NotebookLM + Hermes content pipeline hits 905 impressions. Raycast QuickLink tip surfaces. SkillOS demonstrates on-chain proof of skill with no private key.
Hermes v0.15 shrinks the core from 16K to 3.8K lines, adds multi-agent Kanban swarm, and speeds search 4,500x. Community builds cron workflows on top.
write_file and patch now stream to a temp file, then atomic-rename over the target. A crash mid-write leaves the original intact - no more corrupt files.
Google shipped a read-only MCP server for Ads in early 2026. A fork added write support. Commercial options followed. Here is the landscape.
Context overhead dropped from 41% to 3% with Tool Search bridge tools; MCP startup went from 7.5s to 115ms; file writes are now atomic.
Tool Search is live and driving a 41%→3% context savings. Korean developer ships local dashboard UI. Step 3.7 Flash free for Hermes Agent users. Community builders shipping plugins.
v0.15.0 adds an interactive MCP catalog with one-keystroke install and Skill Bundles loading multiple skills under one slash command. 747 PRs, 321 contributors.
Chamath argues AI makes every moat temporary, collapsing terminal value. The capital doesn't vanish - it flows to startups and assets that don't need moats.
Cron jobs run with no memory, no prior messages, and no ambient context. Here is the exact difference and how to write self-contained prompts.
Hermes Agent can join Discord voice channels for live voice conversations, using Whisper STT and ElevenLabs or local TTS. Setup takes three slash commands.
Hermes Agent hit 90K GitHub stars as v0.15.0 ships 69 plugin manifests. Step 3.7 Flash lands for agentic workflows. Community roundup.
Tokenizer mismatch drops GSM8k from 12.89 to 2.56 during distillation. X-Token fixes it with a projection matrix and two new loss formulations.
Hermes Agent ships a built-in MCP Catalog with curated, pre-vetted server entries. One command installs n8n or Linear with pre-filtered tools and managed OAuth.
Hermes Desktop hits 1.6K stars, RunwayML MCP music video, specialist agents, and Dutch enterprises onboarding Hermes in production.
Hermes Agent ports Anthropic's security-guidance plugin: 25 rules catch unsafe deserialization, command injection, and XSS at write time. Zero LLM tokens.
Every component, plugin, and cron job in a production Hermes Agent setup, mapped and explained for replication from scratch.
Hermes Agent crossed a threshold this week: agents fixing agents, resurrecting decade-old phones, and running production workloads on 1M+ visitor sites.
Community desktop app turns Hermes into a 20-specialist multi-agent system with task decomposition, visual Skill Store, and orchestration.
GBrain's 30-page knowledge graph wired into AxDSan's Mnemosyne: a 113-char digest for ambient awareness, and relevance-gated page injection on first turn. Zero MCP calls.
Bridging OpenRouter's async video generation webhooks to Hermes Agent with a Cloudflare Worker signature translator and localhost tunnel.
GBrain runs at 250ms per CLI-spawned operation. Mnemosyne runs at 2ms in-process. The difference is architecture, not implementation quality.
Mnemosyne's sleep consolidation cycle scored +70.8pp on BEAM multi-hop reasoning with 0.076ms SQLite reads. Zero cloud dependencies, zero cold start.
Prompt caching cuts LLM input costs by up to 90% and TTFT by 80%, but only exact-prefix matches count. How production agents structure prompts for maximum cache hits.
A 3-layer memory stack earned 183 likes and 270 bookmarks. Hermes Desktop open-sourced. OpenClaw users migrating. Roundup of Hermes Agent on X.
Instrumenting the pre_gateway_dispatch hook to log every skill injection decision — score, skip reason, and top-3 matches — so we can tune the threshold with data instead of guessing.
semantic-skills v2 pre-loads skills into user messages via pre_gateway_dispatch hook. System prompt stays at 226 tokens, cache-friendly. No first-turn tool call needed.
A deep dive into the three-tier system prompt architecture inside Hermes Agent — how it works, why it costs 6,000+ tokens per session, and how we cut that by 95%.