Step-by-step build instructions (Design the Charter First, Define a Message Contract, etc.) are implementation content, not design content. Belongs in the repository README. Removing it tightens scope and removes the conflation of design principles with build instructions.

Introduced multi-agent concepts that belong in the Agent Networks document. Purpose — pointing to future work — is now served by the series index. Reformulated as an Exploring entry in the series index.

Three improvement paths (accumulated memory, Charter evolution, model improvement) are either implicit in existing sections or operational concerns beyond this document's scope. The autonomous self-improvement open question moved to the series index Open Questions section.

Final order: Overview → The Agent → The Agentic Loop → Agent Surfaces → Key Benefits → Authentication → Agent Management → Considerations & Trade-offs → Summary → Influences & References. Influences & References closes the document cleanly as the last substantive section.

Two tiers: direct influences (OpenClaw/SOUL.md by Carl Vellotti for Identity & Soul; ReAct, Yao et al. 2022; Constitutional AI, Bai et al. 2022; MCP, Anthropic 2024; Chain-of-Thought; Reflexion; Lilian Weng's survey) and background lineage (cognitive architecture, Unix philosophy, DDD). Framework influences: LangGraph, Semantic Kernel, OpenAI Assistants API, CrewAI, AutoGPT.

Distinguishes three improvement paths with different owners: accumulated memory (agent-internal, continuous), Charter evolution (human-mediated, versioned, observability-driven), and model improvement (separate infrastructure discipline). Autonomous self-improvement flagged as an open design question with safety implications, deliberately out of scope.

Part numbering removed from headers and body text. Series references updated to point to the series index rather than a fixed "Part 2." Standalone Series Index document introduced. Changelog policy: main document shows only the latest entry; full history here.

Replaced base_tools list and discovery_enabled flag with trusted_sources (which MCP servers, APIs, and registries are authorized) and tool_schema (validation contract). Action-class prohibitions moved to guardrails. User-scoped credentials driven by action class, not tool name.

The Charter no longer prescribes a fixed strategy (sliding window, summarization threshold). It defines available infrastructure. The agent manages its own context during reasoning — prescribing a fixed policy is the same mistake as pre-selecting a reasoning strategy.

Removed enabled_types restriction. The Charter defines what memory infrastructure exists; the agent decides which types to use. Restricting memory access is the same mistake as restricting reasoning strategies. Five types remain as a library, not an allowlist.

active_skills field removed from the Standard Message Contract entirely. Skill activation is agent-internal. Surfaces send a message; they have no awareness of the skill library.

Memory retrieval, RAG, and skill activation moved out of Perceive into the Reason stage. Perceive's job is now narrow: normalize input, resolve identity, resolve credentials.

No longer "Defined in the Charter; executed in the Agentic Loop." Now: a library of reasoning patterns the agent draws from. New patterns added to the library are immediately available without Charter changes.

Stage intro: "strategy configured in the Charter" removed. Replaced with: the agent draws from the library, selecting and combining patterns as judgment dictates. Reasoning phase component renamed from "strategy-configured" to "agent-directed." The agent may combine patterns or reason in ways that don't map to any named pattern.

Because the agent may use multiple patterns in one Reason stage invocation, per-strategy units are not a consistent measure. Tokens are the single consistent measure across all approaches.

"Inner loop steps (ReAct)" renamed to "Reasoning phase (strategy-configured)." Description rewritten to cover all strategy classes: CoT (tokens), ReAct (steps), Reflection (revision cycles), Plan-and-Execute (plan steps), Self-Consistency (sample count). Key principle added: the Reason stage interface is strategy-agnostic; what runs inside it is not.

Budget now expressed in strategy-appropriate terms and applies to any strategy. Reasoning trace policy renamed from "Thought trace policy." Updated to clarify trace structure varies by strategy while the retention policy applies uniformly.

Previous framing listed parallel activities (intent understanding, memory retrieval, RAG, planning, tool selection, response generation). Replaced with four defined components: entry conditions (context package Perceive must deliver), inner loop steps (ReAct Thought-Action-Observation cycles), exit conditions (four typed exits: success, step-limit, error, guardrail), and output contract (structured response to Act). The key principle: the Reason stage is not what the LLM does in there — it is a loop with defined interfaces on both sides.

Separate from the outer iteration budget. Governs ReAct steps within a single Reason stage invocation. Follows the Charter-ceiling rule; surfaces may only reduce it. When exhausted, agent exits with best available answer.

Explicit Charter-level decision on whether the ReAct thought trace is logged persistently, retained ephemerally, or discarded. Has cost and compliance implications. Must be defined in the Charter, not left as an infrastructure default.

Three layers: (1) Surface-to-agent inbound — adapter resolves user identity; Charter never handles raw tokens. (2) Agent-to-tools outbound — service-scoped (agent acts as itself) vs. user-scoped credentials (agent acts as the user, OAuth tokens resolved at Perceive). The choice is a per-tool Charter decision. (3) Agent identity — A2A authentication for multi-agent networks; full specification deferred to future document.

Four surfaces with distinct owners: (1) Charter as behavioral source of truth — version-controlled like code; behavioral diffs (how behavior changes, not just which fields) introduced as a first-class requirement. (2) API gateway as operational control plane. (3) Observability platform as feedback loop driving Charter evolution. (4) Skill registry as capability lifecycle — requires explicit governance design.

Version numbering revised from v1.x to v0.x throughout the document. v0.10 is reserved for the finalized document. Previous versions retroactively: v0.01 (April 26), v0.02 (April 29), v0.03 (April 30).

Reframed from "complete governing specification" to "base specification." The Charter establishes the floor: who the agent is, what it does by default, what it must never do. Capability grows from here. "Fully defined, fully capable" replaced with "built on a strong base and designed to grow."

Tool registration reframed from a closed pre-enumerated list to a base set plus a discovery pattern. Tools discovered at runtime via MCP must conform to the schema. Design for discovery, not pre-registration.

Guardrails cover only true hard stops: regulatory prohibitions, irreversible actions, explicit out-of-scope behaviors. A strong Identity & Soul does more safety work than a long rule list because it generalizes to situations no rule anticipated.

active_skills in the Standard Message Contract changed from "list of skill identifiers to activate" (which placed routing logic at the surface) to an optional hint — skill activation based on user intent determined in the Reason stage.

Corrected: "surface configuration drives this routing decision" was wrong. The Charter defines the routing policy; the surface provides the flag. Model selection belongs to the Charter so upgrades require one Charter change, not edits across every surface.

Charter ceilings cannot be exceeded by surface flags. Iteration and timeout budgets: surfaces may only reduce them; more restrictive wins when they conflict. Human-in-the-loop checkpoint corrected from Completion Conditions to Interruption only. Observe stage: logging (async, infrastructure) and context capture (sync, agent-internal) distinguished as separate operations with different consistency requirements.