Field guide / primary domain
228
sources in this field
Updated July 16, 2026
Current thesis
The shortest path to orientation.
The AI agent ecosystem has matured into a recognizable production stack: Pipecat (voice), browser-use (web), Mem0 (memory), Composio (1,000+-app OAuth), RAGFlow/Dify (retrieval), and Mastra (TypeScript-first, 1.77M npm downloads). Cost architecture dominates—80% of tasks are janitorial, and hierarchical model routing (80/15/5 distribution) yields ~10x cost reduction. Configuration converges on three files: SOUL.md (constitution), USER.md (user model), AGENTS.md (playbook). MCP has become survival-level for integration; vendors without MCPs become unusable in agent workflows. Orchestration consensus shifted from single powerful agents to coordinator teams managing sub-agents—swarms fail from coordination, not intelligence. Concrete surfaces shipped: Hermes v0.12.0's Kanban, gstack's 6-specialist Claude teams, @conductor_build's Opus-planning/GPT-5.5-review workflow, NovaStation's unified command center. Finance leads multi-agent debate as the dominant high-stakes pattern. Automation implications are counterintuitive: Marcus Moretti runs Spiral at Every solo (PM+code+support) replacing 60% of old PM work; the unit shifted from job to cross-functional process. Skills distribution is a documentation problem—"skills as markdown," GitHub-backed marketplaces, Browserbase's researched catalogs. Interaction matures into its own discipline: conversation-native UX, cognitive debt, AI Interaction Atlas, and the deepest framing: intelligence as social process composing human-AI systems.
Evidence board
Claims worth carrying forward01Leaner system prompts improve GPT-5.6 performance: internal coding-agent evals show 10–15% score gain, 41–66% token reduction, and 33–67% cost reduction. Trim repeated rules, no-op examples, and unrelated tools while keeping outcome, success criteria, constraints, and tool-routing rules.
02GPT-5.6 migrates better with an outcome-first structure: state what good looks like, add stopping conditions, and use ALWAYS/NEVER only for true invariants. Decision rules outperform blanket absolutes for judgment calls like when to search, ask, or keep iterating.
03GPT-5.6 is more concise by default than GPT-5.5; broad brevity instructions ('Be concise') may now be unnecessary or counterproductive. Use the text.verbosity parameter (low/medium/high) for consistent cross-request control, reserving the prompt for task-specific length requirements.
04Define autonomy and approval boundaries explicitly for agentic GPT-5.6 tasks: name safe local actions (read files, inspect logs, run tests), require confirmation only for external writes, destructive actions, or scope expansion. Repeating 'ask first' causes unnecessary approval pauses.
05Programmatic Tool Calling (PTC) fits bounded record-reduction stages—filtering, joining, deduplication, batching—not simply parallel or dependent calls. Use direct calls when results require semantic judgment, citations, approval, or when each result may change the next decision.
06When migrating to GPT-5.6, preserve current reasoning effort first, run evals before touching the prompt, then remove scaffolding incrementally. Do not rewrite the full prompt stack at once—you cannot isolate whether regressions came from the model, reasoning setting, prompt, tools, or runtime.
07GPT-5.6 prompt migration checklist: switch model → run evals → trim obsolete scaffolding → make one surgical fix per measured regression → re-run evals. Debug regressions on small real traces; identify the failure mode, find the conflicting instruction, edit, rerun.
08For grounded answers, absence of evidence should not automatically become a factual 'no.' Cite only retrieved sources, label inference separately from supported facts, state conflicts between sources, and narrow the answer or report missing evidence instead of guessing.
Adjacent fields
Latest evidence
Adam.GPT
OpenAI's official prompting guide for GPT-5.6 Sol recommends trimming redundant instructions and tools, framing prompts as outcome definitions with stopping conditions, and defining explicit autonomy boundaries for agentic tasks. Internal evals show leaner system prompts improved scores 10–15% while cutting tokens 41–66% and cost 33–67%.
Jason Zook
Jason Zook shares a two-model software development workflow using Claude Opus 4.7 for planning and building, GPT-5.5 for adversarial review, and Playwright for automated UX/UI testing. The loop runs plan→review→update→build→review→fix→deploy, with ~$400/month replacing what feels like a full dev team.
Vasu-Devs
JustHireMe is an open-source, AGPL-licensed desktop app (Tauri + React + Python/FastAPI) that addresses broken job search UX with a local-first, explainable AI pipeline: scrape → quality gate → rank → match (GraphRAG + vector) → generate tailored application materials. All data stays on-device by default, LLM usage is keyless-capable via Ollama/Claude Code CLI, and the architecture prioritizes human-in-the-loop control over blind automation.
Charlie Hills
gstack is Garry Tan's open-source Claude Code setup that transforms a solo AI assistant into a structured 6-role virtual engineering team (CEO, Eng Manager, Designer, QA, Release Engineer, Doc Engineer) via 23 slash-command skills. Tan claims ~810× his 2013 productivity rate measured in normalized logical lines of code, shipping 3 production services and 40+ features in 60 days part-time. The system enforces a think→plan→build→review→test→ship→reflect sprint discipline where each skill feeds context into the next, and supports 10-15 parallel sprints via Conductor.
Ben Tossell
Ben Tossell cancelled Superhuman and built a custom email client using Codex. It mirrors Superhuman's UX patterns but is fully customisable, runs on Gmail CLI, and is agent-native. The core AI logic is minimal: reading labels/archives and updating Gmail filters.
tobi lutke
Tobi Lütke describes River, Shopify's internal AI coding agent living in Slack, designed with one hard constraint: it only works in public channels, not DMs. This forces all AI-assisted work into the open, creating a 'Lehrwerkstatt' (teaching workshop) where knowledge spreads osmotically. In 30 days, 5,938 employees used River across 4,450 channels; it opened 1,870 PRs in one week (~1 in 8 merged PRs). Merge rate climbed from 36% to 77% over two months purely from collective human feedback, not model changes.
elvis
A workflow pattern combining LLM Wikis (structured knowledge stores) with interactive HTML Artifacts creates a bidirectional agent-UI layer. Artifacts are built on top of wikis, can invoke agents, and agents can update artifacts—enabling inbox-zero automation, research scheduling, topic discovery, and live figure generation from a single HTML file.
Ole Lehmann
Ole Lehmann ranks 13 integrations for the Hermes personal agent, with Firecrawl (clean web data, fewer tokens), Browserbase (real browser sessions), Google Workspace, GitHub, and Obsidian as the essential five covering ~80% of needs. Composio handles one-click setup.
JUMPERZ
A two-layer agent workflow: Discord handles plain-English commands and task visibility (especially mobile), while Hermes Kanban is the execution system that assigns, tracks, runs, logs, and proves work. An intake bridge connects both, creating Discord commands → Hermes tasks → Discord task board mirroring.
Shann³
Comprehensive operator guide for Hermes Agent by Nous Research—an open-source autonomous agent framework with 150K GitHub stars and #1 OpenRouter usage. Covers architecture (brain/personality/skillset), four-level deployment progression (laptop to full VPS fleet), multi-agent control room patterns, a 21-step SEO pipeline, and the prototype→production methodology for marketing automation.