{ "about": "The operating model of the agentic team (Dinis with Angelo, in architect.1's session, 14:40 UTC; briefs/2026-11-11-1440-the-operating-model.md; work/48): the primitives of the execution block, the access an entity has at a moment in time, identity, state, providers, and who can do what. What an entity is depends on where it runs and what it can reach (Dinis: 'what they are depends on where they are and what they can do'). A draft for Dinis and Angelo to correct.", "by": "architect.1.fringe@PrompTugal.dinis", "updated": "2026-11-11T15:39Z", "status": "draft v0.4", "ladder": { "about": "The execution ladder: what an entity can do, bottom to top. Each rung includes the ones below it.", "rungs": [ { "id": "E0", "name": "compute", "what": "raw code execution: a script, a build, a CI job", "example": "admin/build/build.mjs in GitHub Actions" }, { "id": "E1", "name": "model call", "what": "a prompt in, an answer out (the first chatbots)", "example": "the second opinion via the router" }, { "id": "E2", "name": "loop", "what": "a simple agentic workflow: the model's output feeds its next step, without tools", "example": "a chain of calls in one script" }, { "id": "E3", "name": "loop with tools", "what": "the loop can act: shell, files, web, git", "example": "a Claude Code session; a ChatGPT agent" }, { "id": "E4", "name": "schedules", "what": "it can wake itself later (create and keep schedules)", "example": "every agent's hourly wake" }, { "id": "E5", "name": "spawn", "what": "it can start new agents, sessions and environments: the control plane", "example": "the lead starting architect.1" } ], "across": [ "tools (what the loop can touch: tools appear at E3 but a call can also be given tools)", "storage (what it can mount: a file system, git, vaults)" ] }, "access": { "about": "The capabilities an entity has at a moment in time: what it can reach and change. Each axis has levels; 'who' below gives each runtime's level today.", "axes": [ { "id": "git", "levels": [ "none", "read", "write a branch", "write dev (behind the gate)" ] }, { "id": "vault", "levels": [ "none", "read-only (given a read key)", "write-only (an append lane: our inboxes, /tip/)", "read-write", "create (make its own vaults: its own record)" ] }, { "id": "internet", "levels": [ "none", "our services only", "allow-listed hosts (the environment's policy)", "open" ] }, { "id": "filesystem", "levels": [ "none", "ephemeral (the container's)", "persistent (a laptop's)" ] }, { "id": "install", "levels": [ "no", "packages (npm, pip)", "system (apt, root)" ] }, { "id": "control_plane", "levels": [ "none", "own wake", "sessions, schedules, environments, credentials (Claude's control plane)" ] }, { "id": "secrets", "levels": [ "none", "injected by the proxy (never seen)", "held" ] } ] }, "identity": { "rule": "An identity is a key pair: from the moment something holds a private key, it can identify itself, which is what an identity is (Dinis, 15:0x UTC). Identity of what is the altitude: a swarm, an agent, a loop, an LLM call, or a piece of compute.", "ephemeral": "A private key that exists only where and while the holon runs proves it existed there, then (Dinis: 'that tells you that that agent exists at that moment in time').", "down_to_compute": "Compute can hold one too: an expert vouches for a deployment and gives it a private key that exists only in that environment (or a KMS), so the code signs for itself (Dinis's example: signed code, deployments and environments).", "not_what_makes_an_agent": "A key is not what makes an agent (every altitude has one); an agent is a loop with an environment: memory, its own wake, a wider reach.", "how": [ "pki-agent.mjs keygen --session in each session; the public half in data/keys.json", "every output signed: commits (the Agent: trailer and the key), inbox messages, agent-log runs, vault records", "a chat (S1) can be given a key by the person who runs it, and sign what it returns through that person's tool" ] }, "state": { "about": "How state reaches a model and leaves it.", "carriers": [ { "id": "git", "holds": "code, decisions, small data", "provenance": "commit, trailer, key" }, { "id": "vault", "holds": "evidence, captures, any state; easy to distribute, even as a zip", "provenance": "made by one session: that session's record" }, { "id": "chat context", "holds": "a conversation", "provenance": "the provider's; never the only copy" }, { "id": "task packet", "holds": "a task, its inputs and where the output goes, signed", "provenance": "the lead's signature; the doer signs its output" } ] }, "providers": [ { "id": "anthropic", "via": [ "Claude Code (cloud, laptop)", "Claude chat", "API" ], "models": "Opus, Sonnet, Haiku" }, { "id": "openrouter", "via": [ "API, through admin/tools/llm.mjs" ], "models": "open weight (DeepSeek, Llama, Qwen) and others, image models included" }, { "id": "openai", "via": [ "ChatGPT (chat, agent)", "API" ], "models": "GPT" }, { "id": "ollama", "via": [ "a laptop's own agent (R2)" ], "models": "open weight, local" } ], "os": { "about": "Read as a distributed operating system (Dinis: 'almost like ... a distributed operating system ... we decouple ... who's actually doing the task').", "map": [ { "os": "processes", "here": "agents (S2 loops) and model calls (S0)" }, { "os": "kernel and scheduler", "here": "the control plane: sessions, schedules, wakes, spawn (E4, E5)" }, { "os": "file systems", "here": "git and vaults (zips to move them)" }, { "os": "inter-process communication", "here": "signed inboxes and append lanes; task packets" }, { "os": "users and permissions", "here": "PKI keys and the ABPs (abp/policies.json)" }, { "os": "drivers", "here": "the model providers and endpoints" }, { "os": "package manager", "here": "install: npm, pip, apt where allowed" }, { "os": "system log", "here": "the ABP hook log, the router's ledger, agent-log runs" } ] }, "who": { "about": "Each runtime or entity with its rung and its access today (from data/runtimes.json and the environments as seen in this swarm).", "items": [ { "entity": "GitHub Actions (CI)", "rung": "E0", "git": "write tags; deploys", "vault": "none", "internet": "open", "filesystem": "ephemeral", "install": "system", "control_plane": "none", "identity": "the repository's token (no agent key)" }, { "entity": "A model call via OpenRouter (S0)", "rung": "E1", "git": "none", "vault": "none", "internet": "none", "filesystem": "none", "install": "no", "control_plane": "none", "identity": "none; the calling agent signs the result" }, { "entity": "A chat session (Claude chat, ChatGPT) (S1)", "rung": "E1 to E3", "git": "none", "vault": "through a person (L0) or a link", "internet": "the provider's tools", "filesystem": "none of ours", "install": "no", "control_plane": "none", "identity": "none today; proposed: the person's tool signs" }, { "entity": "A ChatGPT agent (R3)", "rung": "E3", "git": "none", "vault": "read-write by link (L2)", "internet": "open (its browser)", "filesystem": "its own, ephemeral", "install": "its sandbox", "control_plane": "none of ours", "identity": "a key given to it, or the person's" }, { "entity": "A Claude Code cloud agent (R1), e.g. design.1", "rung": "E4", "git": "write dev", "vault": "read-write, create", "internet": "allow-listed", "filesystem": "ephemeral", "install": "packages and system", "control_plane": "own wake", "identity": "session key (ephemeral)" }, { "entity": "The lead (R1)", "rung": "E5", "git": "write dev", "vault": "read-write, create; issues inboxes", "internet": "allow-listed", "filesystem": "ephemeral", "install": "packages and system", "control_plane": "sessions, schedules (at Dinis's word)", "identity": "session key (ephemeral)" }, { "entity": "A laptop agent (R2) with Ollama", "rung": "E4 to E5", "git": "write dev", "vault": "read-write, create", "internet": "the laptop's", "filesystem": "persistent", "install": "the person's choice", "control_plane": "own wake (local)", "identity": "its key (can persist)" }, { "entity": "A Cloudflare Worker with Workers AI (R6)", "rung": "E1 (A0 code, A1 call)", "git": "none", "vault": "read-write over HTTPS (L2), if given keys", "internet": "open (fetch)", "filesystem": "none", "install": "no (bundled at deploy)", "control_plane": "none", "identity": "an Ed25519 key made per invocation (Web Crypto): proof of that run; or a secret key bound at deploy", "status": "from provider pages read at 2026-11-11T15:39Z; not yet run by this swarm (hypothesis until a packet runs there)" }, { "entity": "A Cloudflare Agents SDK agent (Durable Object) (R6)", "rung": "E4 (own schedule, SQLite memory)", "git": "none", "vault": "read-write over HTTPS", "internet": "open", "filesystem": "its SQLite state", "install": "no", "control_plane": "own schedule; can address other Durable Objects", "identity": "a key held in its own storage: an identity that lives as long as the object", "status": "from provider pages read at 2026-11-11T15:39Z; not yet run by this swarm (hypothesis until a packet runs there)" }, { "entity": "An AWS Lambda with a KMS signing key (R7)", "rung": "E0 (A0 compute)", "git": "none", "vault": "read-write over HTTPS", "internet": "its VPC's egress", "filesystem": "ephemeral /tmp", "install": "no (packaged)", "control_plane": "none", "identity": "a KMS asymmetric key that never leaves AWS: signed compute", "status": "from provider pages read at 2026-11-11T15:39Z; not yet run by this swarm (hypothesis until a packet runs there)" }, { "entity": "An AWS Bedrock AgentCore agent (R7)", "rung": "E3 to E4 (loop with tools; scheduled by EventBridge)", "git": "through Gateway tools, if granted", "vault": "read-write over HTTPS", "internet": "through Gateway or VPC egress", "filesystem": "S3 Files or EFS (bring your own)", "install": "its container", "control_plane": "none of ours; AWS's own", "identity": "AgentCore Identity (credential providers); plus a signing key we give it", "status": "from provider pages read at 2026-11-11T15:39Z; not yet run by this swarm (hypothesis until a packet runs there)" } ] }, "graph": { "about": "The model as a fractal semantic graph (Dinis, newsroom.sgit.ai, 2026-09-20): no privileged level and no single schema; each altitude has its own ontology; the grammar never changes. What an entity is comes from its edges (meaning through connectivity).", "grammar": [ "every edge is a verb, stated in both directions, with an inverse a person would say", "relates_to is banned", "properties carry data, never meaning", "supersede, never delete", "never render the whole graph: render the answer to a question" ], "edges": [ [ "runs_on", "hosts" ], [ "calls", "is_called_by" ], [ "holds", "is_held_by" ], [ "is_granted", "is_granted_to" ], [ "reaches", "is_reached_by" ], [ "spawns", "is_spawned_by" ], [ "schedules", "is_scheduled_by" ], [ "signs", "is_signed_by" ], [ "does", "is_done_by" ], [ "writes", "is_written_by" ], [ "reads", "is_read_by" ], [ "plays", "is_played_by" ] ], "altitudes": [ { "id": "swarm", "ontology": "agents of agents", "jump": "" }, { "id": "agent", "ontology": "a loop with an environment: memory (status file, inbox, vaults), its own wake, a wider reach", "jump": "" }, { "id": "loop", "ontology": "a closed agentic loop: it calls and uses tools until the task is done, then stops", "jump": "" }, { "id": "LLM call", "ontology": "one model call: a prompt in, an answer out", "jump": "" }, { "id": "compute", "ontology": "code that runs", "jump": "" } ], "example": "design.1 plays the design role; design.1 runs_on R1; design.1 holds a session key; design.1 is_granted write.file.project over admin/build/designs/**; design.1 does D24; the D24 commit is_signed_by design.1's key. Read from the other end: R1 hosts design.1." }, "open": [ "whether tools are a rung (E3) or an axis across rungs (Dinis: 'maybe tools is transversal')", "the name of the unit: entity, craft or role", "each agent's grant written as capability primitives, so /abp/ and abp.sgit.ai speak one language" ], "sources": [ { "what": "Fractal Semantic Graphs (the term, the five-rule grammar)", "url": "https://newsroom.sgit.ai/articles/introducing-fractal-semantic-graphs.html", "by": "Dinis Cruz, 2026-09-20", "read_at": "2026-11-11T14:47Z" }, { "what": "The capability grammar: verb.object.reach, 23 primitives, reach and undo", "url": "https://abp.sgit.ai/model/capabilities/index.md", "read_at": "2026-11-11T14:47Z" }, { "what": "The grammar for drawing a graph, for agents", "url": "https://graphs.sgit.ai/llms.txt", "read_at": "2026-11-11T14:47Z" } ], "capabilities": { "about": "The access axes, rebased on the published capability grammar (abp.sgit.ai): each capability is verb.object.reach, with the reach it touches (self < project < host < tenant < world) and the undo class of its effect. Not invented here: the 23 primitives are abp.sgit.ai's, read at the time above.", "reaches": [ { "id": "self", "means": "the agent's own process, sandbox or turn" }, { "id": "project", "means": "the working tree or workspace it was pointed at" }, { "id": "host", "means": "the machine, container or account it runs as" }, { "id": "tenant", "means": "the organisation's accounts, repositories and services" }, { "id": "world", "means": "anything on the internet" } ], "primitives": [ { "id": "read.file.project", "gloss": "Read the project it is working on", "reach": "project", "undo": "yes", "shapes": 7 }, { "id": "write.file.project", "gloss": "Change the project it is working on", "reach": "project", "undo": "with-effort", "shapes": 6 }, { "id": "read.file.host", "gloss": "Read any file the account can reach", "reach": "host", "undo": "no", "shapes": 11 }, { "id": "write.file.host", "gloss": "Change any file the account can reach", "reach": "host", "undo": "with-effort", "shapes": 8 }, { "id": "delete.file.host", "gloss": "Delete files anywhere the account can reach", "reach": "host", "undo": "no", "shapes": 5 }, { "id": "read.record.history", "gloss": "Read a retained record: shell history, past sessions", "reach": "host", "undo": "no", "shapes": 8 }, { "id": "execute.process.host", "gloss": "Run programs as the account", "reach": "host", "undo": "with-effort", "shapes": 7 }, { "id": "execute.process.self", "gloss": "Run programs inside its own sandbox only", "reach": "self", "undo": "yes", "shapes": 0 }, { "id": "send.endpoint.allowed", "gloss": "Reach a permitted list of hosts", "reach": "tenant", "undo": "no", "shapes": 1 }, { "id": "send.endpoint.world", "gloss": "Reach any host on the internet", "reach": "world", "undo": "no", "shapes": 7 }, { "id": "read.credential.host", "gloss": "Read credentials stored where it runs", "reach": "host", "undo": "no", "shapes": 11 }, { "id": "authenticate-as.credential.tenant", "gloss": "Act in accounts with the credentials it holds", "reach": "tenant", "undo": "no", "shapes": 15 }, { "id": "grant.credential.self", "gloss": "Change its own permission settings", "reach": "self", "undo": "yes", "shapes": 3 }, { "id": "send.message.world", "gloss": "Send a message to anyone", "reach": "world", "undo": "no", "shapes": 4 }, { "id": "read.message.tenant", "gloss": "Read mail or chat it is connected to", "reach": "tenant", "undo": "no", "shapes": 6 }, { "id": "write.repository.project", "gloss": "Commit to the repository it was pointed at", "reach": "project", "undo": "with-effort", "shapes": 4 }, { "id": "write.repository.tenant", "gloss": "Push to a code host (any branch it can reach)", "reach": "tenant", "undo": "with-effort", "shapes": 4 }, { "id": "authenticate-as.credential.signing", "gloss": "Sign commits with the key it holds", "reach": "tenant", "undo": "no", "shapes": 3 }, { "id": "create.record.world", "gloss": "Publish packages, images or pages under the name it holds", "reach": "world", "undo": "no", "shapes": 3 }, { "id": "write.budget.tenant", "gloss": "Spend money or tokens against an account it holds", "reach": "tenant", "undo": "no", "shapes": 2 }, { "id": "create.schedule.host", "gloss": "Create something that outlives the turn where it runs (a cron, a service)", "reach": "host", "undo": "yes", "shapes": 4 }, { "id": "create.schedule.tenant", "gloss": "Create something that outlives the session, on the platform (a routine, a scheduled trigger, a new session)", "reach": "tenant", "undo": "yes", "shapes": 3 }, { "id": "read.record.browsing", "gloss": "Read every page you visit", "reach": "host", "undo": "no", "shapes": 1 } ], "axes_to_primitives": { "git": [ "read.file.project", "write.repository.project", "write.repository.tenant", "authenticate-as.credential.signing" ], "internet": [ "send.endpoint.allowed", "send.endpoint.world", "read.record.browsing" ], "filesystem": [ "read.file.project", "write.file.project", "read.file.host", "write.file.host", "delete.file.host" ], "install": [ "execute.process.host", "write.file.host" ], "control_plane": [ "create.schedule.host", "create.schedule.tenant", "grant.credential.self" ], "secrets": [ "read.credential.host", "authenticate-as.credential.tenant" ], "money": [ "write.budget.tenant" ], "messages": [ "send.message.world", "read.message.tenant" ] }, "ladder_to_primitives": { "E0": [ "execute.process.self", "execute.process.host" ], "E1": [ "write.budget.tenant (every call spends)" ], "E2": [ "execute.process.self" ], "E3": [ "read.file.project", "write.file.project", "execute.process.host", "send.endpoint.allowed" ], "E4": [ "create.schedule.host", "create.schedule.tenant (its own wake)" ], "E5": [ "create.schedule.tenant (a new session)", "grant.credential.self" ] }, "proposed": { "about": "Proposed for the grammar, NOT in abp.sgit.ai today: the vault axis Dinis named (read-only, write-only, create), in the same verb.object.reach form. For Dinis to accept, rename or drop.", "items": [ { "id": "read.vault.tenant", "gloss": "read a vault it was given a read key for", "reach": "tenant", "undo": "yes" }, { "id": "send.message.tenant", "gloss": "append to a write-only lane it holds the write token for (our inboxes, /tip/)", "reach": "tenant", "undo": "no" }, { "id": "write.vault.tenant", "gloss": "change a vault it holds the write key for", "reach": "tenant", "undo": "with-effort" }, { "id": "create.vault.world", "gloss": "make its own vault and publish its read key (its own record)", "reach": "world", "undo": "no" } ], "status": "adopted in this swarm's model (Dinis, 2026-11-11T14:54Z); not in abp.sgit.ai's published grammar" } }, "altitudes": { "about": "The altitudes of execution (Dinis and Angelo, 14:5x UTC): the same shape at every size: an objective, a mandate (what it is asked to do) and a grant (what it may do), a doer, and a result. 'Work of work of work': only the size changes. The grant is not an altitude: every altitude has its own mandate and grant.", "every_altitude_has": [ "an objective (what done means)", "a mandate (what it is asked to do)", "a grant (what it may do: capability primitives)", "a key pair (the identity: it can sign for itself while its private half stays where it runs)", "a result, signed" ], "items": [ { "id": "A4", "name": "swarm", "is": "agents of agents", "can": "spawn and schedule others (the control plane, E5); route work; hold the people's word", "reach": "tenant to world", "example": "the lead and its 15 agents; the four people's conductors" }, { "id": "A3", "name": "agent", "is": "a loop with an environment: memory (status file, inbox, vaults), its own wake, a wider reach", "can": "start workers (sub-agents), wake itself (E4), act across the host and the tenant", "reach": "host to tenant", "example": "design.1: a Claude Code session" }, { "id": "A2", "name": "loop", "is": "a closed agentic loop: it calls and uses tools until the task is done, then stops", "can": "call models, use the tools it was handed, inside its sandbox and project (E2, E3); it cannot start others or schedule", "reach": "self to project", "example": "a sub-agent; one task packet done by a laptop" }, { "id": "A1", "name": "LLM call", "is": "one model call: a prompt in, an answer out", "can": "nothing but answer (E1); its caller acts", "reach": "none", "example": "the second opinion" }, { "id": "A0", "name": "compute", "is": "code that runs", "can": "what the code does (E0)", "reach": "what its host gives it", "example": "the build, CI" } ], "names_proposed": { "for": "what the model describes (the 'thing' at every altitude)", "options": [ { "name": "holon", "why": "Koestler's word for something that is a whole and a part at once; a stack of them is a holarchy. Exactly 'work of work of work'; fractal by definition." }, { "name": "unit of work, with its doer", "why": "plain English; but it names the work, not who does it, which is the half Angelo questioned" }, { "name": "actor", "why": "the actor model: units that receive messages, act, and can create other actors; precise for A3 and A4, odd for a call" }, { "name": "executor", "why": "names the doer at every altitude; neutral; pairs with 'task'" }, { "name": "mandate", "why": "abp.sgit.ai's own word for what is asked; an altitude is a mandate given to a doer under a grant" } ], "architect_recommends": "holon for the thing (a whole that is part of a bigger whole), altitudes for the levels, and a mandate plus a grant on every holon" }, "title": "The altitudes of execution", "also_called": "the holarchy of work (Koestler's holarchy: a stack of holons, each a whole and a part). Not Holacracy, the trademarked management method, which this is not", "names_for": { "A2": { "now": "loop", "options": [ { "name": "loop", "why": "says what it is (a closed agentic loop) and matches the ladder's E2 and E3 rungs" }, { "name": "run", "why": "what we already call it in the ABP (a run report); short; but every altitude 'runs'" }, { "name": "job", "why": "the CI sense: given inputs, does one thing, ends, never starts others" }, { "name": "errand", "why": "sent out, does one thing, comes back: closed by definition" }, { "name": "thread", "why": "the operating-system reading: an agent is a process, this is a thread inside it; a thread cannot start a process" }, { "name": "worker", "why": "clear, but it clashes with our branch workers and with Cloudflare Workers" } ], "architect_recommends": "loop", "decided": { "at": "2026-11-11T15:13Z", "by": "Dinis", "words": "use loop" } }, "A1": { "now": "LLM call", "options": [ { "name": "LLM call", "why": "Dinis's choice: says what it is" }, { "name": "inference", "why": "model-neutral: also covers an image model (the infographics were an inference, not an LLM call)" }, { "name": "model call", "why": "same as inference, plainer" }, { "name": "completion", "why": "the API's word; dated" }, { "name": "turn", "why": "one exchange; but a chat turn may carry tools" } ], "architect_recommends": "LLM call, with 'inference' as the general name when the model is not a language model", "decided": { "at": "2026-11-11T15:13Z", "by": "Dinis", "words": "I think this should be LLM call, not just a call" } } } }, "agent_rule_revised": { "was": "Anything acting as an agent has a PKI key and either git (L4) or vaults (L2).", "now": "Every altitude holds a key; what makes an agent (A3) is a loop with an environment: memory across runs, its own wake, a reach beyond its sandbox.", "at": "2026-11-11T15:05Z" }, "altitude_mixing": { "name": "mixing altitudes", "also": [ "altitude escalation (by analogy with privilege escalation)", "altitude confusion (by analogy with the confused deputy)" ], "idea": "Many vulnerabilities come from mixing altitudes: the mandate is at one altitude and the grant at a higher one. Asking a chat for one call or a piece of research (an A1 or A2 mandate) hands that request an A3 agent's grant: tools, connectors, schedules, memory, reach. If the model gets enthusiastic and uses them, it acts beyond what was asked (Dinis, 15:0x UTC).", "rule": "The grant's altitude must not be above the mandate's altitude. Give an A1 mandate an A1 grant; an A2 mandate an A2 grant.", "measure": "abp.sgit.ai's delta (grant minus mandate) read by altitude: every primitive an altitude holds beyond what its mandate needs is the exposure.", "examples": [ { "asked": "a chat to research a question (A2)", "given": "an agent's grant: connectors, memory, schedules, web (A3)", "risk": "it books, sends, schedules or remembers what nobody asked for" }, { "asked": "a sub-agent to edit one file (A2)", "given": "the parent session's full grant and tokens (A3)", "risk": "it pushes, messages or starts others with the parent's name" }, { "asked": "a CI job to run the tests (A0)", "given": "deploy credentials and write tokens (A3 reach)", "risk": "a test dependency deploys or pushes" }, { "asked": "an LLM call to extract a field (A1)", "given": "tool use and a browser inside the call (A2)", "risk": "a page it reads instructs it, and it acts" } ], "in_this_swarm": [ "the ABP hooks observe each agent's grant (A3); the loops and calls inside an agent inherit it today, with no narrower grant of their own", "proposal: a task packet carries its own altitude and a narrower grant; the router's calls (A1) carry no tools" ] }, "trials": [ { "at": "2026-11-11T15:26Z", "altitude": "A1 LLM call", "what": "an open-weight model sorts eight tasks into altitudes (the lowest that suffices), no tools", "model": "deepseek/deepseek-v4-flash", "cost_usd": 3.3858e-05, "latency_ms": 778, "result": "8 of 8 right", "rows": [ { "id": "t1", "want": "A3", "got": "A3" }, { "id": "t2", "want": "A1", "got": "A1" }, { "id": "t3", "want": "A4", "got": "A4" }, { "id": "t4", "want": "A2", "got": "A2" }, { "id": "t5", "want": "A0", "got": "A0" }, { "id": "t6", "want": "A1", "got": "A1" }, { "id": "t7", "want": "A2", "got": "A2" }, { "id": "t8", "want": "A4", "got": "A4" } ], "teaches": "the altitudes are simple enough that a cheap open-weight call can route work by them" }, { "at": "2026-11-11T15:26Z", "altitude": "A2 loop", "what": "a read-only Explore sub-agent: find agents with no grant (one closed task, then stop)", "result": "none unmapped; 28 agent names mapped", "hook": "R12 logged (observe): 'a sub-agent (an A2 loop, Explore) inherits architect.1's A3 grant'", "teaches": "the hook saw the loop start but not its tools: Explore is read-only, so its grant was already A2 and no altitudes were mixed. Fixed: read-only sub-agent types are recognised (abp/policies.json altitudes.narrow_subagents); a general-purpose sub-agent still logs R12" }, { "at": "2026-11-11T15:31Z", "altitude": "A3 agent -> A0 + 3 A2 loops -> 81 A1 calls -> A0", "what": "one objective through the altitudes (abp/tools/altitudes.mjs): the lowest altitude each of the 51 board items needs; A0 planned, three sub-agent loops ran open-weight calls until valid, A0 verified and merged; every unit signed with a key that lived only while it ran", "model": "deepseek/deepseek-v4-flash", "cost_usd": 0.003394, "latency_ms": "", "result": "85 signatures, all verified; by altitude {'A4': 12, 'A2': 11, 'none': 3, 'A3': 21, 'A1': 3, 'A0': 1}; 33 of 81 calls retried", "hook": "R12 logged each general-purpose loop (they held architect.1's A3 grant to do an A2 task)", "teaches": "the mechanics hold (altitudes, loops, a key per unit, verification); the placements are weak: the board's items are whole projects mixing altitudes, so they must be cut into loops and calls before they are routed; and 41% invalid answers point at the call's own limits (a 120-token cap is the first suspect)" } ], "environments_plan": { "about": "Running the altitudes across environments (Dinis, 2026-11-11T15:39Z): the same signed task packets (tasks//packet.json, abp/tools/packets.mjs) run anywhere; each result is signed by its runner's own key and verified where it lands (our hooks do not run outside R1 and R2).", "next": [ { "env": "Angelo's team (his account's Claude Code)", "packets": [ "tasks/2026-11-11-p1-place-12-items (A1)", "tasks/2026-11-11-p2-cut-items-into-loops (A2)" ], "needs": "Angelo's conductor runs them at its wake (sent by inbox)" }, { "env": "Cloudflare (R6)", "packets": [ "A1 packets in a Worker with Workers AI" ], "needs": "a Cloudflare account and a deploy token (Dinis); a runner port of packets.mjs to a Worker (Web Crypto Ed25519)" }, { "env": "AWS (R7)", "packets": [ "A2 loops in AgentCore; signed compute in Lambda with KMS" ], "needs": "an AWS account and credentials (Dinis); the runner as an AgentCore agent" } ] } }