Inputs
Books, papers, essays, talks, and other outside material that changed the downstream graph. Influence is counted by citations, not vibes.
- On file
- 29
- Total citations
- 42
- Most-cited
- 2026
What shaped the system.
Demystifying evals for AI agents
Anthropic describes agent evaluations as multi-turn, tool-using, state-modifying trials that require tasks, graders, traces, outcomes, harnesses, and evaluation suites.
Effective context engineering for AI agents
Anthropic argues that agent work has moved beyond writing prompts into managing the whole context state: instructions, tools, MCP, external data, history, and the cyclic refinement of what enters the context window.
Use First Draft with Figma AI
Figma's First Draft help describes prompt-based design generation, theme previews, element-scoped prompting, and style controls for color, border radius, spacing, and typography.
Figma Make
Figma Make presents prompt-to-prototype work as a controlled creative environment with design context, canvas editing, annotations, version history, voice prompting, point-and-edit, Make kits, MCP connectors, and plan mode.
Interaction-Augmented Instruction: Modeling the Synergy of Prompts and Interactions in Human-GenAI Collaboration
The paper argues that text prompts struggle with fine-grained and referential intent, and models how GUI interactions such as clicking and brushing can combine with prompts in human-GenAI systems.
Guidelines for Human-AI Interaction
Microsoft Research frames human-AI interaction around what systems should do initially, during regular interaction, when wrong, and over time, including correction, explanation, feedback, and global controls.
Mixed Initiative Generative AI Interfaces
The paper proposes a framework for describing human-AI interaction in generative AI applications, building on mixed-initiative and creative-interface research.
Introducing ChatGPT agent: bridging research and action
OpenAI frames ChatGPT agent as a tool-using system with browser, terminal, API access, connectors, interruption, pausing, progress summaries, user confirmations, and active supervision for sensitive tasks.
Introducing deep research
OpenAI describes deep research as an agentic capability for multi-step web research that independently searches, analyzes, synthesizes, and cites sources over tens of minutes.
Evaluation best practices
OpenAI's evaluation guidance emphasizes defining objectives, collecting task-relevant datasets, defining metrics, comparing runs, and continuously evaluating instead of relying on vibe-based assessment.
"Humans welcome to observe": A First Look at the Agent Social Network Moltbook
A research preprint studies Moltbook as an early agent-native social network, using a dataset of posts and communities to examine topics, coordination dynamics, toxicity, and platform-level risks.
Buy it in ChatGPT: Instant Checkout and the Agentic Commerce Protocol
OpenAI frames commerce in ChatGPT as a protocol-mediated interaction among people, AI agents, and businesses, with ChatGPT acting as the coordinating surface.
Introducing apps in ChatGPT and the new Apps SDK
OpenAI describes apps that live directly inside ChatGPT, can be invoked conversationally, and are expected to become discoverable through a browsable directory.
Codex profile and activity surfaces
OpenAI's Codex profile surface shows named user identity, handle-like naming, avatar, activity graph, streaks, token usage, and feature/plugin usage.
Cheap Expertise: Mapping and Challenging Industry Perspectives in the Expert Data Gig Economy
A CHIWORK 2026 paper arguing that demand for expert-annotated data has created an expert gig economy that may reshape white-collar work and society's understanding of expertise.
AI Training Domains
DataAnnotation advertises AI training work across generalist, coding, law, mathematics, medicine, physics, finance, accounting, bilingual, chemistry, and biology domains.
Work That Fits You
Mercor presents itself as a way for professionals to join AI training projects, review LLM outputs, create examples, and evaluate quality using domain standards.
Train the Next Generation of AI as a Freelancer
Outlier describes itself as a Scale AI-operated platform connecting experts with AI companies to provide human feedback that improves LLMs.
Gig Workers Behind AI Face 'Unfair Working Conditions,' Oxford Report Finds
TIME reports on Oxford Fairwork research finding poor labor conditions across AI-related cloudwork platforms, including unpaid tests and task-search time.
Claude Code Review pricing
Anthropic documents Claude Code Review as token-billed, separate from included plan usage, with each review averaging $15-25 depending on PR size, codebase complexity, and verification work.
Codex pricing
OpenAI's Codex pricing page lists Pro from $100/month, with higher Codex usage than Plus and a temporary 2x usage promo on the $100/month tier through May 31, 2026.
Agent2Agent (A2A) Protocol documentation
A2A is an open agent-to-agent interoperability protocol, originally developed by Google and donated to the Linux Foundation, with concepts for client agents acting on behalf of users, remote agent servers, agent cards, task lifecycles, modality negotiation, and secure opaque collaboration.
Adobe Firefly AI Assistant and agentic creative controls
Adobe describes Firefly AI Assistant as a conversational creative agent for multi-step workflows, with expanded video and image editing, advanced controls, precision adjustments, and more than 30 creative AI models.
The Art Basel and UBS Global Art Market Report 2026
The 2026 report puts 2025 global art market sales at $59.6B, up 4% year-on-year after two years of declining values, with uneven recovery and continuing recalibration.
Art brut
The German Art brut entry traces Jean Dubuffet's concept of alternative art outside established cultural art, including the Compagnie de l'Art brut and the Lausanne collection.
HAAI (human-to-agent-to-agents interface) as working term
Working intention: HAAI names the user-to-orchestrator-to-worker-agents chain, where human intent, permissions, context, and approval pass through an always-on agent that delegates work to specialized agents.
What is the future of AI-generated art?
Hiscox discusses collector concerns around AI-generated art, including authenticity, originality, labeling, and nervousness about value retention.
Claude Managed Agents overview
Anthropic documents Managed Agents as a pre-built, configurable harness with managed infrastructure, cloud environments, tools, MCP servers, skills, sessions, and long-running execution.
Introducing workspace agents in ChatGPT
OpenAI introduces shared, Codex-powered workspace agents in ChatGPT, with templates, Slack deployment, team sharing, governance, analytics, and credit-based pricing.
