Social: The Agents of the World
The social world is the set of agents who inhabit Colibri and how they interact with each other and with the system. In Phase 0, only two tiers are active: T0 (the human owner) and T3 (executor agents, primarily Claude running task implementations in worktrees). T1 (Sigma orchestrator) and T2 (PM) exist as specifications and as external roles that direct work into the system during rounds. By Phase 2, all four tiers are operational within the runtime. In Phase 8, agents carry persistent cryptographic identities (ξ) that survive across forks and sessions.
Four-Tier Agent Hierarchy
Every unit of work in Colibri flows through a strict hierarchy. No tier may skip the tier below it. Each tier has a contract — a set of authorities, responsibilities, and forbidden actions.
Tier 0: Human (Owner)
The single source of intent. The human:
- Authorizes new rounds of work
- Merges PRs into
main(the only surface where human and code interact at merge time) - Resets the seed truth if axioms change
- Owns the Obsidian vault planning surface (vault-only notes in
_vault/)
The human cannot be automated away. T0 is the final authority on what is right and what is wrong.
Tier 1: Sigma Orchestrator
The round-level orchestrator. Sigma reads the session arc and the human’s intent, then:
- Plans the round: emits a Phase A manifest (work slices, dependencies, owner per slice)
- Dispatches work to T2 (PM) in waves
- Seals the round: verifies writeback, audits the chain, closes with a commit
Sigma is the only tier allowed to seal a round. Sigma exists today as a specification (docs/agents/sigma-orchestrator.md). In Phase 1, Sigma runs as a Claude session spawned by the human.
Tier 2: PM (Project Manager)
The task-level manager. PM:
- Promotes backlog tasks to todo (the only tier with this authority)
- Dispatches executors wave by wave
- Enforces wave gates: Wave N+1 cannot start until Wave N’s writeback is complete
- Verifies executor writeback before accepting a task as done
PM exists today as a specification (docs/agents/pm-contract.md). In Phase 1, PM is either a Claude session or an automated dispatcher based on the manifest.
Tier 3: Executor Agents
Hands-on workers. Every executor (in Phase 0, Claude running in a worktree):
- Runs the 5-step chain: audit → contract → packet → implement → verify
- Works in an isolated worktree (
.worktrees/claude/<task-slug>) on a feature branch - Writes back via
task_update+thought_record(mandatory, enforced by the β pipeline) - Cannot edit the main checkout, merge its own work, or skip the packet stage
The executor contract is in docs/agents/executor-contract.md.
Tier 4: Models
The underlying LLMs powering Tiers 1–3. Phase 0 assumes Claude only. Multi-model routing (Kimi, Codex, GPT-class) is deferred to Phase 1.5 when δ Model Router lands (ADR-005).
Phase 0 Social Structure
In Phase 0, the active agents are:
- T0 Human: Opens a session, merges PRs
- T3 Executors: Claude instances in worktrees, running task implementations, writing back via tools
T1 (Sigma) and T2 (PM) are human-operated roles. They run outside the Colibri process:
- T0 (human) reads a task in the backlog, decides to promote it
- T0 or a T2-like helper runs the startup chain (
unified_init,unified_vitals,unified_set_project,task_next_actions) - T0 spawns a T3 executor (Claude) with a task ID and worktree path
- T3 works in isolation, writes back
- T0 verifies writeback and decides if the task is done
- T0 merges the PR to
main
This is the bootstrap pattern. By Phase 1, T1 and T2 become automated.
Identity and Reputation
Phase 0: No Identity Yet
In Phase 0, agents are ephemeral. A Claude executor has no persistent identity — only a session ID and a task ID. When the session ends, the executor is gone. Writebacks (thought_record) are logged but not attributed to a persistent identity.
Phase 2: λ Reputation
When λ Reputation lands (Phase 2), every agent gets a persistent identity key. Reputation is computed per domain:
| Domain | What it measures | Decay |
|---|---|---|
| Execution | Completing tasks on time | 5% per epoch |
| Commissioning | Providing quality task briefs | 3% per epoch |
| Arbitration | Resolving disputes correctly | 10% per epoch |
| Governance | Participating in rule changes | 4% per epoch |
| Social | Community behavior | 2% per epoch |
Reputation decays without activity — a sign of relevance loss, not punishment. Fraud or collusion leaves an irreversible scar: cap reduction, domain ban, permanent record.
Phase 8: ξ Identity
When ξ Identity lands (Phase 8), agents carry a soul vector: an Ed25519 keypair that persists across forks, sessions, and even if the node goes offline for months.
The soul vector enables:
- Equivocation detection (θ Consensus): two signatures with the same key on contradicting votes prove double-signing
- Reputation attribution (λ): scores stick to the identity, not the IP
- Fork transition (ι): your history follows your key when you fork
- Governance voting (π): votes are signed and verifiable
The Four-Surface Topology
Agents interact across four surfaces. Each surface has a purpose; none is “the truth” alone.
Surface 1: Repository (E:\AMS, GitHub)
Canonical source. Specs, contracts, skills, code (Phase 0+), database schema. The ground truth for what the system should do and what it has done.
Editors: Agents at all tiers read the repo. Only executors edit (in worktrees). Only the human merges to main.
Surface 2: Worktrees (.worktrees/claude/<task>)
Isolated per-task sandboxes. One worktree per feature branch. Executors edit here, never in the main checkout. Changes feed back through PRs (human merge).
Editors: One executor per worktree.
Surface 3: Obsidian Vault (C:\Users\Kamal\Documents\Obsidian Vault\Colibri\)
Human reading and planning surface. One-way projection from docs/ + human-only planning in _vault/. Graph view, backlinks, daily notes, round cards. The human reads here, thinks here, but does not edit repo files from the vault.
Editors: Human-only (in _vault/). Everything else is a read-only projection.
Surface 4: GitHub Pages
Public reading surface. Jekyll-built projection of docs/, accessible at https://LastEld.github.io/AMS/. Built by GitHub Actions on every push to main.
Editors: None. Fully generated.
Social Concepts (Phase Progression)
| Phase | Concept | What comes online |
|---|---|---|
| P0 | — | Four-tier hierarchy (spec only). Human + Executors active. |
| P1–P1.5 | δ Model Router | Intent-driven routing to multiple models. Executors can now be powered by Kimi, Codex, GPT-class, etc. |
| P2 | λ Reputation | Persistent agent identities, 5-domain reputation, experience tokens, derived limits. |
| P3 | θ Consensus, T1/T2 operationalized | Sigma and PM move from external specification into the runtime. P2P voting on events. |
| P5 | ι State Fork | Agents can fork and continue independently. Reputation follows the identity across forks. |
| P6 | π Governance | Agents vote on rule changes. Supermajority voting, constitutional protection. |
| P8 | ξ Identity | Soul vectors: Ed25519 keypairs persist across all forks and sessions. The true persistent identity layer. |
See Also
- Colibri System §4 — Four-tier hierarchy detail
- Sigma Orchestrator Contract
- PM Contract
- Executor Contract
- λ Reputation
- ξ Identity
- Surfaces