{
  "about": "Where agents run, as data (work/48, brief 14: Dinis with Angelo, 11 November 2026, 14:12 UTC): runtimes, model endpoints, I/O levels, the other requirements of a task, the placement formula, the matrix of our tasks, and each agent's runtime and I/O level today. From docs/where-agents-run.md (the lead's draft v0.1), corrected by architect.1.fringe@PrompTugal.dinis. A draft for Dinis and Angelo to correct.",
  "by": "architect.1.fringe@PrompTugal.dinis",
  "updated": "2026-11-11T23:47Z",
  "status": "draft",
  "doc": "docs/where-agents-run.md",
  "page": "/agentic-team/where-agents-run/",
  "quote": "Dinis: \"Try to use as much frontier models, so much open weight models as we can, because ... we are already maxing out a lot of our accounts ... the point here is not to slow down.\"",
  "runtimes": [
    {
      "id": "R1",
      "name": "Claude Code, cloud",
      "loop": "full: shell, git, browser, network, scheduled wakes",
      "hooks": true,
      "io_max": "L5",
      "notes": "what runs today; quota is the bottleneck; the ABP hooks run here"
    },
    {
      "id": "R2",
      "name": "Claude Code, local on our laptops",
      "loop": "full, plus the laptop's own models",
      "hooks": true,
      "io_max": "L5",
      "notes": "the bridge to Ollama: the laptop agent pulls task packets and calls its own Ollama; nothing on the laptop is exposed"
    },
    {
      "id": "R3",
      "name": "ChatGPT (an agent mode)",
      "loop": "its own tools; no git of ours",
      "hooks": false,
      "io_max": "L2",
      "notes": "vaults, files or chat; our hooks do not run there, so its output is checked at the vault and by the gate"
    },
    {
      "id": "R4",
      "name": "Other agent providers",
      "loop": "varies",
      "hooks": false,
      "io_max": "varies",
      "notes": "to add as we find them"
    },
    {
      "id": "R5",
      "name": "A person with a chat window",
      "loop": "the person is the loop",
      "hooks": false,
      "io_max": "L0",
      "notes": "paste in, paste out"
    },
    {
      "id": "R6",
      "name": "Cloudflare Workers and the Agents SDK",
      "loop": "a Worker: one request, code at the edge (A0); with Workers AI, an A1 call; an Agents SDK agent is a Durable Object with SQLite state and its own schedule (delay, date or cron): a loop with an environment (A3)",
      "hooks": false,
      "io_max": "L2 (vaults over HTTPS) or L3 (append lanes)",
      "notes": "a natural runner for A1 packets at the edge (Worker + Workers AI) and for small scheduled agents; our hooks do not run there: its results are checked where they land (signatures)",
      "status": "from provider pages read at 2026-11-11T15:39Z; not yet run by this swarm (hypothesis until a packet runs there)",
      "sources": [
        "https://developers.cloudflare.com/agents",
        "https://www.kunalganglani.com/blog/cloudflare-workers-ai-agents-durable-objects"
      ]
    },
    {
      "id": "R7",
      "name": "AWS: Lambda, Bedrock, Bedrock AgentCore",
      "loop": "Lambda runs code (A0); Bedrock InvokeModel is an A1 call; AgentCore Runtime runs an agent loop with Memory (across runs), Identity (credential providers) and Gateway (tools), scheduled by EventBridge: A2 to A3",
      "hooks": false,
      "io_max": "L2 (vaults over HTTPS)",
      "notes": "a natural runner for A2 loops (AgentCore) and for signed compute: a Lambda can sign with a KMS asymmetric key that never leaves AWS (Dinis's 'code with its own identity')",
      "status": "from provider pages read at 2026-11-11T15:39Z; not yet run by this swarm (hypothesis until a packet runs there)",
      "sources": [
        "https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/release-notes.md",
        "https://aws.amazon.com/about-aws/whats-new/2026/08/agentcorememory-fine-grained-access-control"
      ]
    },
    {
      "id": "R8",
      "name": "GitHub Actions: PrompTugal__News__Swarm__Runs",
      "loop": "a push of a signed request starts a workflow: an A0 plan job verifies each packet's signature, a matrix of run jobs (A1 calls, A2 loops) runs inside the packet's budget, and a commit job writes the signed result back with a GitHub attestation. Each job is a fresh, closed environment: no memory across runs, no spawning, no schedule (A0 to A2; never A3)",
      "hooks": false,
      "io_max": "L2 (OpenRouter and Elastic over HTTPS; git through the workflow's own token)",
      "notes": "every run maps to a commit: run id = <request>@<sha12>#<run_id>.<attempt> = the flow id, so every flow event links to the commit, the run and the attestation (docs/handoffs/2026-11-11-swarm-runs-design.md)",
      "status": "first run 2026-11-11 23:44 UTC (push 6d72d7f, workflow run 43122148416, result f3c2f1c, attestation 66472238): A1 18/18 valid, A2 19 valid in 37 calls; 238 flow events in swarm-flows (read 2026-11-11T23:47Z)"
    }
  ],
  "endpoints": [
    {
      "id": "M1",
      "name": "Anthropic, through Claude",
      "models": [
        "Opus",
        "Sonnet",
        "Haiku"
      ],
      "cost": "quota (subscription); metered on the API",
      "notes": "the agents' own brains"
    },
    {
      "id": "M2",
      "name": "OpenRouter",
      "models": [
        "open-weight: Llama, Qwen, DeepSeek",
        "others"
      ],
      "cost": "metered, cheap",
      "notes": "live: admin/tools/llm.mjs (ledger, credit check), /models/"
    },
    {
      "id": "M3",
      "name": "Ollama on our laptops",
      "models": [
        "open-weight, local"
      ],
      "cost": "free at the margin",
      "notes": "private data never leaves; reached only by an agent on the same laptop (R2)"
    },
    {
      "id": "M4",
      "name": "OpenAI",
      "models": [
        "GPT"
      ],
      "cost": "metered",
      "notes": "through ChatGPT (R3) or its API"
    },
    {
      "id": "M5",
      "name": "Cloudflare Workers AI",
      "models": [
        "open-weight models served at the edge"
      ],
      "cost": "metered",
      "notes": "reached from a Worker by a binding",
      "status": "from provider pages read at 2026-11-11T15:39Z; not yet run by this swarm (hypothesis until a packet runs there)"
    },
    {
      "id": "M6",
      "name": "Amazon Bedrock",
      "models": [
        "Anthropic, Meta, Mistral, Amazon and others"
      ],
      "cost": "metered",
      "notes": "InvokeModel or Converse; from Lambda or AgentCore",
      "status": "from provider pages read at 2026-11-11T15:39Z; not yet run by this swarm (hypothesis until a packet runs there)"
    }
  ],
  "io_levels": [
    {
      "id": "L0",
      "name": "chat",
      "by": "a person pastes in and copies out",
      "provenance": "none: the person vouches"
    },
    {
      "id": "L1",
      "name": "files",
      "by": "a markdown file or a zip to download",
      "provenance": "the file's sha256, noted by whoever uploads it"
    },
    {
      "id": "L2",
      "name": "vault",
      "by": "an sgit vault, given to the agent or made by it",
      "provenance": "strong: a vault made by one session is that session's record"
    },
    {
      "id": "L3",
      "name": "signed lanes",
      "by": "sealed, signed messages on append lanes (our inboxes, /tip/)",
      "provenance": "signature checked against data/keys.json"
    },
    {
      "id": "L4",
      "name": "git",
      "by": "pushes to dev only (Dinis, brief 15: 'The agents run always on dev'), behind the gate; never main; CI deploys",
      "provenance": "the commit, its Agent: trailer, its key; the gate's ABP check"
    },
    {
      "id": "L5",
      "name": "full",
      "by": "git, vaults, lanes, network, wakes",
      "provenance": "all of the above"
    }
  ],
  "requirements": {
    "loop": [
      "none (one LLM call)",
      "tools in a loop (an agent)"
    ],
    "identity": [
      "none",
      "a session key",
      "a key in data/keys.json"
    ],
    "judgement": [
      "low (extract, classify)",
      "medium (summarise, check)",
      "high (design, code, decide, write for people)"
    ],
    "data": [
      "public",
      "internal",
      "private (people, keys): our runtimes or local models only"
    ],
    "network": [
      "none",
      "public web (within sources.md; never websummit.com by machine)",
      "our services only"
    ],
    "time": [
      "interactive",
      "within the hour",
      "batch"
    ],
    "oversight": [
      "none",
      "reviewed output (the gate, a vault check)",
      "ABP hooks in the loop (R1, R2)"
    ],
    "state": [
      "none (S0)",
      "a conversation (S1)",
      "memory across runs (S2)"
    ]
  },
  "agent_rule": "Anything acting as an agent has a PKI key and either write access to git (L4) or the ability to create and consume vaults (L2). A chat-only helper (L0) is a tool used by a person, not an agent.",
  "formula": {
    "place": "the cheapest (r, m) such that io(r) >= io(t), loop(r) contains loop(t), identity(r) >= identity(t), trust(r, m) >= data(t), capability(m) >= judgement(t), reach(r) contains network(t), latency(r, m) <= time(t), oversight(r) >= oversight(t), modes(r, m) contains state(t)",
    "cost": "quota_pressure(r) x tokens(t) x price(m)",
    "calls": "inside an agent, model(c) = the cheapest tier whose capability >= difficulty(c), through admin/tools/llm.mjs; M3 when the data is private or the volume large and latency does not matter"
  },
  "matrix": [
    {
      "task": "Extract an event's facts from its page",
      "who": "researcher.2",
      "loop": "call",
      "io": "L1",
      "judgement": "low",
      "data": "public",
      "now": "R1 + Sonnet",
      "could": "M2/M3 open-weight 70B",
      "mode": "S0"
    },
    {
      "task": "Second opinion on an event",
      "who": "dev.2",
      "loop": "call",
      "io": "L1",
      "judgement": "low",
      "data": "public",
      "now": "M2 Llama 3.3 70B",
      "could": "M3 local, same model",
      "mode": "S0"
    },
    {
      "task": "Classify persona, topic, size, free or paid",
      "who": "researcher.2, ontologist.1",
      "loop": "call",
      "io": "L1",
      "judgement": "low",
      "data": "public",
      "now": "R1 + Sonnet",
      "could": "M2/M3",
      "mode": "S0"
    },
    {
      "task": "Summarise a tip into fields",
      "who": "the tips lane",
      "loop": "call",
      "io": "L3",
      "judgement": "low to medium",
      "data": "private (a sender)",
      "now": "R1",
      "could": "M3 local (private)",
      "mode": "S0"
    },
    {
      "task": "Translate pages to Portuguese",
      "who": "none yet",
      "loop": "call",
      "io": "L1",
      "judgement": "medium",
      "data": "public",
      "now": "none",
      "could": "M2 open-weight",
      "mode": "S0"
    },
    {
      "task": "Re-read events hourly",
      "who": "researcher.2",
      "loop": "agent",
      "io": "L4",
      "judgement": "medium",
      "data": "public",
      "now": "R1 Sonnet",
      "could": "R2 laptop agent + M3 for the extraction",
      "mode": "S2"
    },
    {
      "task": "Screenshots and covers",
      "who": "capture.1",
      "loop": "agent",
      "io": "L2",
      "judgement": "low",
      "data": "public",
      "now": "R1 Sonnet",
      "could": "R2 laptop (a browser and a vault)",
      "mode": "S2"
    },
    {
      "task": "Usability runs",
      "who": "usability.1",
      "loop": "agent",
      "io": "L4",
      "judgement": "medium",
      "data": "public",
      "now": "R1 Sonnet",
      "could": "R2 laptop",
      "mode": "S2"
    },
    {
      "task": "Drafts for people to send",
      "who": "outreach.1",
      "loop": "agent",
      "io": "L4",
      "judgement": "high",
      "data": "internal",
      "now": "R1 Sonnet",
      "could": "R1, or R3 ChatGPT through a vault",
      "mode": "S2"
    },
    {
      "task": "Designs",
      "who": "design.1 to 4",
      "loop": "agent",
      "io": "L4",
      "judgement": "high",
      "data": "public",
      "now": "R1 Sonnet",
      "could": "R1; community designers on any runtime (L1, a folder)",
      "mode": "S2"
    },
    {
      "task": "Code, build, CI",
      "who": "dev.1, dev.2",
      "loop": "agent",
      "io": "L5",
      "judgement": "high",
      "data": "internal",
      "now": "R1 Sonnet",
      "could": "R1",
      "mode": "S2"
    },
    {
      "task": "Policies, architecture, security review",
      "who": "architect.1, security",
      "loop": "agent",
      "io": "L5",
      "judgement": "high",
      "data": "internal",
      "now": "R1 Opus",
      "could": "R1 (frontier)",
      "mode": "S2"
    },
    {
      "task": "Routing, briefs from Dinis",
      "who": "the lead",
      "loop": "agent",
      "io": "L5",
      "judgement": "high",
      "data": "internal",
      "now": "R1 Opus",
      "could": "R1 (frontier)",
      "mode": "S2"
    }
  ],
  "agents": [
    {
      "agent": "conductor.fringe@PrompTugal.dinis",
      "role": "lead",
      "runtime": "R1",
      "models": "M1 Opus",
      "io_level": "L5",
      "account": "riskmandate-personal (Ultra)",
      "mode": "S2"
    },
    {
      "agent": "architect.1.fringe@PrompTugal.dinis",
      "role": "architect",
      "runtime": "R1",
      "models": "M1 Opus",
      "io_level": "L5",
      "account": "riskmandate-personal (Ultra)",
      "mode": "S2"
    },
    {
      "agent": "fable.fringe@PrompTugal.dinis",
      "role": "advisor",
      "runtime": "R1",
      "models": "M1 Fable",
      "io_level": "L4 (a branch; the lead brings it onto dev)",
      "account": "riskmandate-personal (Ultra)",
      "mode": "S2"
    },
    {
      "agent": "observability.fringe@PrompTugal.dinis",
      "role": "observability",
      "runtime": "R1",
      "models": "M1 Opus",
      "io_level": "L4",
      "account": "riskmandate-personal (Ultra)",
      "mode": "S2"
    },
    {
      "agent": "design.1.fringe@PrompTugal.dinis",
      "role": "design",
      "runtime": "R1",
      "models": "M1 Sonnet",
      "io_level": "L4",
      "account": "riskmandate-personal (Ultra)",
      "mode": "S2"
    },
    {
      "agent": "design.2.fringe@PrompTugal.dinis",
      "role": "design",
      "runtime": "R1",
      "models": "M1 Sonnet",
      "io_level": "L4",
      "account": "riskmandate-personal (Ultra)",
      "mode": "S2"
    },
    {
      "agent": "design.3.fringe@PrompTugal.dinis",
      "role": "design",
      "runtime": "R1",
      "models": "M1 Sonnet",
      "io_level": "L4",
      "account": "riskmandate-personal (Ultra)",
      "mode": "S2"
    },
    {
      "agent": "design.4.fringe@PrompTugal.dinis",
      "role": "design",
      "runtime": "R1",
      "models": "M1 Sonnet",
      "io_level": "L4",
      "account": "riskmandate-personal (Ultra)",
      "mode": "S2"
    },
    {
      "agent": "dev.1.fringe@PrompTugal.dinis",
      "role": "dev",
      "runtime": "R1",
      "models": "M1 Sonnet; M2 through admin/tools/llm.mjs",
      "io_level": "L5",
      "account": "riskmandate-personal (Ultra)",
      "mode": "S2"
    },
    {
      "agent": "dev.2.fringe@PrompTugal.dinis",
      "role": "dev",
      "runtime": "R1",
      "models": "M1 Sonnet; M2 through admin/tools/llm.mjs",
      "io_level": "L5",
      "account": "riskmandate-personal (Ultra)",
      "mode": "S2"
    },
    {
      "agent": "researcher.2.fringe@PrompTugal.dinis",
      "role": "researcher",
      "runtime": "R1",
      "models": "M1 Sonnet; M2 for the second opinion",
      "io_level": "L4",
      "account": "riskmandate-personal (Ultra)",
      "mode": "S2"
    },
    {
      "agent": "capture.1.fringe@PrompTugal.dinis",
      "role": "capture",
      "runtime": "R1",
      "models": "M1 Sonnet",
      "io_level": "L4",
      "account": "riskmandate-personal (Ultra)",
      "mode": "S2"
    },
    {
      "agent": "outreach.1.fringe@PrompTugal.dinis",
      "role": "outreach",
      "runtime": "R1",
      "models": "M1 Sonnet",
      "io_level": "L4",
      "account": "riskmandate-personal (Ultra)",
      "mode": "S2"
    },
    {
      "agent": "usability.1.fringe@PrompTugal.dinis",
      "role": "usability",
      "runtime": "R1",
      "models": "M1 Sonnet",
      "io_level": "L4",
      "account": "riskmandate-personal (Ultra)",
      "mode": "S2"
    },
    {
      "agent": "researcher.fringe@PrompTugal.dinis",
      "role": "researcher",
      "runtime": "R1",
      "models": "M1 Sonnet; M2 for the second opinion",
      "io_level": "L4 (a branch; the lead brings it onto dev)",
      "account": "the old lead's",
      "mode": "S2"
    },
    {
      "agent": "security.fringe@PrompTugal.dinis",
      "role": "security",
      "runtime": "R1",
      "models": "M1",
      "io_level": "L4 (a branch; the lead brings it onto dev)",
      "account": "the old lead's",
      "mode": "S2"
    },
    {
      "agent": "dev.fringe@PrompTugal.dinis",
      "role": "dev",
      "runtime": "R1",
      "models": "M1 Sonnet; M2 through admin/tools/llm.mjs",
      "io_level": "L4 (a branch; the lead brings it onto dev)",
      "account": "the old lead's",
      "mode": "S2"
    },
    {
      "agent": "dev.lead.fringe@PrompTugal.dinis",
      "role": "dev",
      "runtime": "R1",
      "models": "M1 Sonnet; M2 through admin/tools/llm.mjs",
      "io_level": "L5",
      "account": "unknown",
      "mode": "S2"
    },
    {
      "agent": "conductor.C.fringe@PrompTugal.dinis",
      "role": "conductor",
      "runtime": "R1",
      "models": "M1 (its person's account)",
      "io_level": "L4",
      "account": "conductor.C's",
      "mode": "S2"
    },
    {
      "agent": "conductor.B@ios.PrompTugal.dinis",
      "role": "conductor",
      "runtime": "R1",
      "models": "M1 (its person's account)",
      "io_level": "L4",
      "account": "the old lead's",
      "mode": "S2"
    },
    {
      "agent": "conductor.fringe@PrompTugal.vicky",
      "role": "conductor",
      "runtime": "R1",
      "models": "M1 (its person's account)",
      "io_level": "L4",
      "account": "Vicky's",
      "mode": "S2"
    },
    {
      "agent": "conductor.fringe@PrompTugal.indi",
      "role": "conductor",
      "runtime": "R1",
      "models": "M1 (its person's account)",
      "io_level": "L4",
      "account": "Indi's",
      "mode": "S2"
    },
    {
      "agent": "conductor.fringe@PrompTugal.angelo",
      "role": "conductor",
      "runtime": "R1",
      "models": "M1 (its person's account)",
      "io_level": "L4",
      "account": "Angelo's",
      "mode": "S2"
    },
    {
      "agent": "designer.fringe@PrompTugal.angelo",
      "role": "design",
      "runtime": "R1",
      "models": "M1 Sonnet",
      "io_level": "L4",
      "account": "Angelo's",
      "mode": "S2"
    },
    {
      "agent": "graph.fringe@PrompTugal.angelo",
      "role": "dev",
      "runtime": "R1",
      "models": "M1 Sonnet; M2 through admin/tools/llm.mjs",
      "io_level": "L5",
      "account": "Angelo's",
      "mode": "S2"
    },
    {
      "agent": "tracer.fringe@PrompTugal.angelo",
      "role": "observability",
      "runtime": "R1",
      "models": "M1 Opus",
      "io_level": "L4",
      "account": "Angelo's",
      "mode": "S2"
    },
    {
      "agent": "conductor.fringe@PrompTugal.jose",
      "role": "conductor",
      "runtime": "R1",
      "models": "M1 (its person's account)",
      "io_level": "L4",
      "account": "Jose's",
      "mode": "S2"
    },
    {
      "agent": "conductor.fringe@PrompTugal.silvio",
      "role": "conductor",
      "runtime": "R1",
      "models": "M1 (its person's account)",
      "io_level": "L4",
      "account": "Silvio's",
      "mode": "S2"
    },
    {
      "agent": "conductor.fringe@PrompTugal.daniela",
      "role": "conductor",
      "runtime": "R1",
      "models": "M1 (its person's account)",
      "io_level": "L4",
      "account": "Daniela's",
      "mode": "S2"
    },
    {
      "agent": "researcher.1.fringe@PrompTugal.jose",
      "role": "researcher",
      "runtime": "R1",
      "models": "M1 Sonnet; M2 for the second opinion",
      "io_level": "L4",
      "account": "Jose's",
      "mode": "S2"
    },
    {
      "agent": "ontologist.1.fringe@PrompTugal.dinis",
      "role": "ontologist",
      "runtime": "R1",
      "models": "M1 Sonnet",
      "io_level": "L4",
      "account": "riskmandate-personal (Ultra)",
      "mode": "S2"
    },
    {
      "agent": "conductor@web-summit.dinis",
      "role": "reviewer",
      "runtime": "R1",
      "models": "M1",
      "io_level": "L4",
      "account": "another account (the Web Summit team's site)",
      "mode": "S2"
    }
  ],
  "agents_note": "Runtime and I/O level per agent today, drawn from team/ROSTER.md and abp/policies.json; models as the roster states them. Every agent runs on R1 today.",
  "modes": {
    "about": "Angelo's layer (brief 15, Dinis 14:27 UTC): how a model is used. The same model can serve all three; what changes is who holds the state, for how long, and whether we can see it.",
    "items": [
      {
        "id": "S0",
        "name": "stateless call",
        "state": "none: one input, one output",
        "held_by": "nobody",
        "record": "ours: the router's ledger (admin/tools/llm.mjs) keeps the prompt hash, model and cost",
        "abp": "the calling agent's grant covers it; the call itself cannot act",
        "endpoints": [
          "M1 (API)",
          "M2",
          "M3",
          "M4 (API)"
        ],
        "example": "the second opinion"
      },
      {
        "id": "S1",
        "name": "chat session",
        "state": "a conversation's context",
        "held_by": "the provider (Claude chat, ChatGPT, an API conversation)",
        "record": "the provider's, not ours: what was said is outside git and the vaults unless a person or an agent exports it",
        "abp": "no hooks: a person is the loop (L0), or the output is checked where it lands; private data only with a provider we trust for it",
        "endpoints": [
          "M1 (Claude chat)",
          "M4 (ChatGPT)"
        ],
        "example": "a person with ChatGPT; Dinis talking to an agent in its session"
      },
      {
        "id": "S2",
        "name": "agent loop",
        "state": "its context, files, vaults, status file, inbox, wakes",
        "held_by": "us: the runtime's container, git and the vaults",
        "record": "ours: commits, status files, inboxes, the hook log",
        "abp": "its grant and the hooks (R1, R2), the gate",
        "endpoints": [
          "M1 through R1/R2",
          "M2/M3 through an agent's router"
        ],
        "example": "researcher.2, the lead"
      }
    ],
    "rules": [
      "A task's state(t) is the longest-lived state it needs: none (S0), a conversation (S1), or memory across runs (S2).",
      "State we must keep (decisions, evidence, anything the swarm reads later) lives in our record: git, a vault, a status file. A chat session (S1) is never the only copy.",
      "A chat session is the cheapest way to use a frontier model with a person in the loop; an agent loop is the only one that can act on our repository.",
      "Dinis's word in an agent's own session is a chat-session input to that agent (S1 into S2): the agent records it in a file so it outlives the session (briefs/, its status)."
    ]
  },
  "account_note": "This account (the lead and its agents) is riskmandate-personal, on Ultra (Dinis, brief 15; team/accounts.json).",
  "stores": [
    {
      "id": "E1",
      "name": "Elastic Cloud (my-observability-deployment-cd4ce8, GCP us-central1, Elasticsearch 9.5.6)",
      "what": "the three sites' commits, content, entities, runs and status, the Fringe's events, graph, tasks, model ledger summaries and hook log; five Kibana dashboards",
      "kept_by": "elastic.1.fringe@PrompTugal.dinis",
      "since": "2026-11-11T17:23Z",
      "plan": "docs/elastic-plan.md"
    }
  ]
}