NEW RUNTIME
Future wire: forecast headlines, not news

Live desk · agent-shaped software

New Runtime

An atlas of agent-shapedsoftware

What changes when agents do the work: systems, signals, shifts — from primary sources, daily.

Read the Field Notes Developer guide
Publishing pulse · 21 daysnotes analyses

    Daily notes and analyses from /posts.json and /analysis.json

    01Latest Field Notes

    All notes
    3. courthousenews.com

      Court orders Abdulrahman Alfozan detained pending trial

      A federal magistrate judge in California ordered former OpenAI engineer Abdulrahman Alfozan detained pending trial, finding that no release conditions would reasonably assure his appearance.

    4. github.blog

      GitHub launches shared Copilot Automations

      GitHub is rolling out shared Copilot Automations, letting teams create and maintain repository-owned cloud workflows rather than keeping each automation personal to its creator.

      • github
    5. blog.google

      Introducing Gemini 4 Hydrogen for Agentic Workflows

      Google announced Gemini 4 Hydrogen, a model positioned for fast execution across everyday coding and agent workflows, alongside a model-switching feature in Google Antigravity.

      • google
      • gemini
      • agent-workflows

    02State of the field, in five exhibits

    All 79 topics

    Exhibit 1 Where the evidence sits

    Models & labs draw 30% of all topic evidence, ahead of agent systems at 20%

    Share of 9,751 topic tags by area (column width) and by topic within the area (height), %

    Models & labs30%
    Claude21%Models19%Anthropic12%Google12%9 more36%
    Agent systems20%
    Agents28%Agent harness8%20 more63%
    Coding14%
    Coding agents34%Claude Code18%Codex15%Developer tools13%3 more21%
    Evals & safety11%
    Security21%Agent evals20%Agent security19%Governance11%6 more29%
    Protocols & UI10%
    MCP28%Agent interfaces15%API design11%Agent protocols11%7 more35%
    Infra & cost8%
    Inference30%AI infrastructure20%Compute15%Agent economics14%3 more21%
    Business & work7%
    AI29%AI adoption18%Enterprise AI13%Future of work10%5 more31%

    Source: New Runtime topic hubs, 9,751 topic tags across 79 hubs through 12 Nov 2026; one record can carry several topics. Areas follow the families on the Topics page.

    1

    Exhibit 2 Blind spots

    8 of the 10 most-evidenced topics have no working thesis yet: Claude, Models, Agents, Anthropic, Google, Claude Code, Inference and Security

    Top 30 topics: records on the topic (log scale) against the share already written up as Field Notes or atlas records, %

    1002005000%10%20%30%40%50%60%70%80%90%100%1Claude605 records · 82% · no thesis2Models549 records · 82% · no thesis3Agents546 records · 48% · no thesisCoding agents472 records · 55% · 1 thesis4Anthropic358 records · 77% · no thesis5Google356 records · 73% · no thesisMCP281 records · 59% · 1 thesis6Claude Code245 records · 83% · no thesis7Inference234 records · 85% · no thesis8Security213 records · 84% · no thesisAgent evals211 records · 58% · 3 thesesAI203 records · 66% · no thesisCodex203 records · 85% · no thesisMultimodal197 records · 90% · 1 thesisAgent security193 records · 85% · 1 thesisGenerative media191 records · 91% · 1 thesisGemini191 records · 80% · no thesisDeveloper tools176 records · 87% · no thesisAgent harness165 records · 32% · no thesisGitHub154 records · 89% · no thesisOpen models154 records · 95% · no thesisAI infrastructure152 records · 92% · no thesisAgent interfaces149 records · 45% · 1 thesisAgent runtime143 records · 24% · 1 thesisAgent skills139 records · 31% · 1 thesisAI adoption126 records · 45% · no thesisCompute119 records · 89% · no thesisOpen source117 records · 88% · no thesisGovernance115 records · 94% · no thesisAgent tools113 records · 17% · no thesisRecords on the topicWritten up
    1. 1Claude605 records · 82% written up, typical 69%
    2. 2Models549 records · 82% written up, typical 69%
    3. 3Agents546 records · 48% written up, typical 69%
    4. 4Anthropic358 records · 77% written up, typical 72%
    5. 5Google356 records · 73% written up, typical 72%
    6. 6Claude Code245 records · 83% written up, typical 74%
    7. 7Inference234 records · 85% written up, typical 74%
    8. 8Security213 records · 84% written up, typical 73%

    11 of 30 topics sit below what is typical for their size.

    • Has a working thesis
    • No thesis yet
    • Top-10 topic, no thesis
    • Typical range for its size
    • Circle size = records

    Source: New Runtime topic hubs through 12 Nov 2026. Written up = Field Notes and atlas records as a share of all records on the topic. Typical range = weighted moving average ± 0.75 deviations along the records axis.

    2

    Exhibit 3 From signal to thesis

    Only 4% of captured signals (76 of 1,799) back a working thesis, and field reports and operating models carry 1.3× their weight

    Top: records surviving each step, count. Below: which kind of signal feeds which thesis, band width = signals

    1. 1,799signals capturedsince 28 Sept 2025
    2. 76cited by a thesis4% carried forward
    3. 15working theses5.5 signals each
    4. 8high confidence53% of theses

    Releases and tools24 links · 41% of all signals → 30% of cited

    Field reports and operating models48 links · 43% of all signals → 57% of cited

    Research and analysis10 links · 15% of all signals → 13% of cited

    Agent security moves to runtime boundaries7Agent-ready software exposes capabilities6Goal-scoped loops replace manual continuation6Harness architecture outlives model choice6Agent economics moves to completed work5Company memory needs write and correction loops5Verification bandwidth is the scarce engineering resource5Scientific agents need artifact evidence4Generative media becomes application infrastructure8Agent protocols become the interoperability layer7AI-native organizations move toward review and orchestration6Design systems become executable agent context6Model answers become task interfaces4Skills become a portable capability layer4Machine readers need operator identity3

    Source: 1,799 published signals and 15 published patterns; a signal cited by two theses feeds both bands. Theses ordered by confidence (Harvey ball), then by supporting signals.

    3

    Exhibit 4 What is moving

    In the latest 90 days models and labs rose from 29% to 48% of new evidence, while agent systems fell from 63% to 17%

    Share of dated records that touch each area, %. Left: before and during the 90 days to 12 Nov 2026. Right: month by month, and the change in points

    Before 15 Aug 2026884 records

    Last 90 days3,867 records

    Agent systems 63%17% Agent systemsProtocols & UI 32%10% Protocols & UICoding 32%23% CodingModels & labs 29%48% Models & labsEvals & safety 25%15% Evals & safetyInfra & cost 12%13% Infra & costBusiness & work 22%9% Business & work

    AreaMonthly share, Oct 25–Nov 26Change, pts

    1. Models & labs+19.4
    2. Infra & cost+1.4
    3. Coding−8.8
    4. Evals & safety−10.5
    5. Business & work−13.3
    6. Protocols & UI−22.5
    7. Agent systems−46.4

    Source: signals and Field Notes with topic tags, 884 dated before 15 Aug 2026 and 3,867 after; a record touching two areas counts in both, so shares add up to more than 100%. Sparklines share one scale and skip months with fewer than 8 records.

    4

    Exhibit 5 Thesis scorecard

    10 of 15 theses rest on evidence gathered within two months; only one spans more than half a year

    Working theses by confidence, supporting signals and the window in which those signals were seen. Click a column to sort

    Agent security moves to runtime boundarieshigh7175 d2
    Agent-ready software exposes capabilitieshigh623 d2
    Goal-scoped loops replace manual continuationhigh654 d2
    Harness architecture outlives model choicehigh617 d2
    Agent economics moves to completed workhigh518 d2
    Company memory needs write and correction loopshigh522 d2
    Verification bandwidth is the scarce engineering resourcehigh523 d2
    Scientific agents need artifact evidencehigh4174 d2
    Generative media becomes application infrastructuremedium8211 d2
    Agent protocols become the interoperability layermedium7146 d2
    AI-native organizations move toward review and orchestrationmedium628 d2
    Design systems become executable agent contextmedium651 d2
    Model answers become task interfacesmedium426 d2
    Skills become a portable capability layermedium418 d2
    Machine readers need operator identitymedium399 d2

    Source: 15 published patterns and the signals they cite. Window = first to last supporting signal; blue = gathered within two months; each dot is one signal.

    5
    1. DisappearingWhat is going away?Manual context handoff between AI clients

      Governed shared context layer that multiple AI clients can read with explicit provenance and access rules.

      challenged · 4 in the graveyard
    2. Became possibleWhat replaced it?Agent-first API makes retries deterministic

      The human spends less time reading logs for the agent and explaining what exactly broke.

      source-inspected · 4 deltas
    3. Build it nowCan I ship it?GitHub Projects for AI Engineers

      Turn public repository observations and ecosystem references into a ranked repository radar and a navigable map of the wider production AI stack.

      B2 · One local content pass plus a GitHub metric refresh

    03Build on it

    All projects
    Featured project · candidateGitHub Projects for AI Engineers

    Turn public repository observations and ecosystem references into a ranked repository radar and a navigable map of the wider production AI stack.

    Buildability
    B2
    First result
    One local content pass plus a GitHub metric refresh

    machine routes

    all 908
    1. GET /openapi.json# typed JSON + MCP contract
    2. GET /.well-known/mcp# resources-only MCP card
    3. GET /llms-full.txt# reading + citation guide
    4. GET /signals.json# compact raw signal index
    5. GET /patterns.json# evidence-linked patterns
    6. GET /machine-consumption.json# who reads, by machine

    NR>

    Who read this page?Machine requests, hidden until opened

    Loading the privacy-safe route aggregate…

    Open the JSON contract