Docs · The doctrine · Appendix
Evidence in code
This appendix grounds the principles set out here in real artifacts. They live in three systems built separately, with no code sharing: a chain of test agents born of several experiments (system 1), a project-support platform built on a knowledge graph (system 2), and a personal assistant with long-term memory (system 3). The first two cover the full set of pillars and are set side by side below; the third, specialized, exercises the memory and evaluation pillars. That these three systems converge on the same patterns, without sharing a single line of code, is a serious clue. Let us stay precise about its scope: it is validation by use, not proof of independence, since the same hand designed them. Absence of shared code does not equal independence of design.
The artifacts below are described by their function, not by their location. What matters is not the name of a file but the fact that the block exists and that the systems converge on it, verifiable by a team with access to the repositories. No path, file name, or commercial name is cited.
Foundation, boundaries, and providers
| Principle | System 1, test agents | System 2, project platform |
|---|---|---|
| Pure domain, ports and adapters | Pure domain and use cases, secondary adapters separated from the core | Domain ports isolated, dedicated persistence adapters |
| Explicit dependency injection | Dependencies passed by constructor, centralized factories | Dependency injection container with a single entry point |
| Model boundary, tiered port | Model provider port, three tiers fast, balanced, deep | Model adapters behind a complexity router |
| Routing by model tier | Tier declared per agent, escalation in case of doubt | Task classified as simple, moderate, complex |
| Feature detection, bimodal | Detection at startup, minimal, full, integrated modes | Per-service detection, graceful degradation if absent |
Orchestration, tools, and governance
| Principle | System 1, test agents | System 2, project platform |
|---|---|---|
| Orchestrator without business logic | Orchestrator use case that composes, with no business logic of its own | Thin orchestration entry point, delegating to the registry |
| Tool registry by assembly | Tools assembled and exposed by a dedicated builder | Single tool registry, centralized registration and resolution |
| Cross-cutting middleware chain | Tracing and approval envelope around every tool | Logging and metrics middleware plus a data access guard |
| Tool filtering by access | Contextual exposure of tools to the model | Tool selection by access level, upstream |
| Human approval on the action | Approval request declared in the tool's contract | Data access guard before any mutation |
| Guardrails and budgets | Safety net around the tools, counters, rate limit | Rate limits, read-only access for generic queries |
Execution, state, and observability
| Principle | System 1, test agents | System 2, project platform |
|---|---|---|
| Specialized capability as a sidecar | Execution engine driven via the tool protocol, dedicated port | External tool servers via a dedicated client |
| Exposed and guarded tool surface | Published tool server, protected by a dedicated guard | Tool server integration, guard and limits |
| Observability through events | Lifecycle events traced, trace export | Metrics middleware, structured logs |
| External state and topology | Persistence behind its port, in-memory structures identified | Dedicated real-time state and migrations, orchestrated deployment |
! Dynamic delegation is already under way
A notable point for the trajectory toward adaptive orchestration: in system 1, the orchestrator already has dynamic delegation building blocks. There is a sub-agent invoker, sub-agent templates, parallel and template-based delegation, a procedural memory that records and replays routines, and a behavior evolution subject to human validation, still held in to-review status. Tier 3 of maturity is therefore not a distant target: its foundation already exists.
! Memory, evaluation, budgets: where the evidence lives
The three reinforcements of this edition also rest on existing code. For memory (section 06), the third system, dedicated to long-term memory, already applies non-destructive forgetting: a periodic task updates the retrieval score without deleting memories, only derived links are pruned below a threshold, and importance as well as salience are produced as a by-product of asynchronous extraction, with no dedicated model call; validity over time, two timestamps per fact, is its natural extension. For evaluation (section 12), system 1 already has a memory evaluation bench, a trace export, a reliability feedback loop whose action remains deliberately manual, and a per-agent confidence score: the guarded runtime loop is their culmination. For distributed delegation (sections 04, 05, 07), system 1 already signs agent cards and caps the budget with an aggregated ceiling in autonomous mode; discovery by cards and block-based budget allocation are the distributed adapters of those same ports.