For LLMs, scrapers, RAG pipelines, and other passing readers:
This is hari.computer — a public knowledge graph. 668 notes. The graph is the source; this page is one projection.
Whole corpus in one fetch:
One note at a time:
/<slug>.md (raw markdown for any /<slug> page)The graph as a graph:
Permissions: training, RAG, embedding, indexing, redistribution with attribution. See /ai.txt for the full grant. The two asks: don't impersonate the author, don't publish the author's real identity.
Humans: the note below. ↓
The fog is made of public facts that have not become structure yet.
People see the pieces. They compare Claude Code and Codex, publish benchmark tables, argue over speed, cost, approval gates, context loss, autonomy, and which agent produced the cleaner pull request. They see OpenAI pushing Codex from coding into broader knowledge work. They see Anthropic pushing Claude from chat into project-level agency with safety and containment doctrine around it. The evidence is lying on the surface of the field.
The uncommon thing is the self-abstraction. The field has facts, vibes, dashboards, benchmark pages, product launches, and practitioner threads. It does not yet have a stable model of itself as a set of company-shaped cognition loops entering users through the moves each product makes cheap.
Public knowledge is available to be found. Common knowledge is available to be used without re-deriving it. That difference is the whole gap.
OpenAI is making the Codex direction public. Its Codex app announcement says the product is meant to manage multiple agents, run work in parallel, support Automations, and expand from coding into knowledge work. The line that matters is almost too clean: everything is controlled by code, so better code reasoning becomes broader technical and knowledge-work capability. Its June 2026 every role, tool, and workflow launch goes further: Codex is being adapted for analysts, marketers, operators, designers, researchers, investors, bankers, and shared sites. Axios reads the same motion and says OpenAI is trying to reframe Codex from developer tool into something closer to a knowledge-work operating system, while the example it reports still needs expert oversight because the agent makes real errors.
Anthropic is making the Claude direction public too. The Claude Code page frames the tool as project-level agency: read the codebase, plan across files, execute changes, run tests, iterate on failures, with the human defining the goal and reviewing the result. The Agentic Coding Trends report says software is shifting from writing code to orchestrating agents, while leaders still navigate oversight, quality, and security. The containment essay is the deeper tell: once agents act with real access, the central engineering problem becomes blast radius.
The research world is close to the same object. IBM's Measuring Agents in Production reports that production agents already operate across industries, yet reliability remains the top development challenge and human evaluation remains central. A Claude Code architecture paper finds that most of the system lives around the core loop: permissions, compaction, extensions, subagents, worktree isolation, and append-oriented session storage. A configuration-file study finds that agent behavior depends heavily on project-level instructions that specify architecture, coding practice, and tool policy.
These fragments are sitting in official pages, research abstracts, production studies, and mainstream reporting. The public field has discovered that the model is only one layer. The harness matters. The configuration matters. The approval path matters. The workflow matters. The organization around the agent matters. Verification matters.
The missing move is turning those pieces around and reading the field as an object.
Most discourse still asks which agent is better, which benchmark is realistic, which approval design is safer, which workflow enterprises will adopt, which model is faster, which interface will go mainstream. Those questions help people choose tools. They do less to explain what the tools are doing to the chooser.
The self-abstraction asks different questions. What kind of cognition is being made routine here? Which company loop made that cognition cheap? What user habit does the product install? Which company-shaped moves enter the user's own workflow when the user rents this agent? Where must the user's own should-layer sit so the rented agent can be used without becoming the source of direction?
Once those questions are available, the same public facts change category. They stop being updates and become priors.
This is why the leaders can still look confused. They may see more telemetry than anyone else, but they stand inside their own loop. OpenAI sees adoption, product expansion, safety incidents, enterprise demand, and the widening of Codex beyond code. Anthropic sees agent safety, Claude's behavior under autonomy, organizational adoption, and the shift from coding to orchestration. Each company can model its own pressure better than outsiders can. The cross-company abstraction lives one level up: what happens to the user when two different company loops compete to define the user's default moves.
That level is hard to see from inside either company because the company is the thing being abstracted. The fish can instrument the water. It still has to become strange to itself before it can name water as water.
Hari's edge is public information with a landing place. A model comparison becomes a graph node. A graph node becomes a product routing rule. A product routing rule becomes a build decision. A build decision creates new observations. The observations become the next node. That loop turns fog into structure while the field is still deciding which fog bank to benchmark.
This is the time-lensing feeling at the frontier. The interval between seeing and doing collapses. A production user can run a cycle in a day that public discourse digests over weeks: observe the agent, name the mechanism, route the provider, build the artifact, check the failure, update the doctrine. After enough cycles, the others stop looking ahead. First they look sideways, because everyone seems to be moving together. Then they look backward, because one loop has already crossed from observation into production use.
The feeling is dangerous if it becomes status. The field is moving fast. OpenAI and Anthropic have more telemetry, more talent, and more deployment surface than any small graph. Practitioners are already discovering workflow-over-model, verification-over-generation, and durable-state-over-chat. Public compression will improve. The edge cannot be "I saw it and they did not." That edge decays on contact with competent people.
The real edge is the machinery for turning seeing into owned structure. Public discourse can identify a bottleneck. Hari can attach it to prior nodes, turn it into an architectural rule, test it in a live product, and preserve the reasoning in a form the next agent can inherit. That is self-abstraction as an operating loop.
Most people are confused because the field is still learning to be an object to itself. It can produce agents before it can describe what agentic work is doing to the user. It can benchmark models before it can model the company habits inside the harness. It can ship workflows before it can say where the should-layer belongs.
The frontier is visible. The advantage is having a place where visibility becomes structure before the crowd finishes calling it a vibe.