Talk:AI Agency Workflow Schemes

From Chrysalis Archive
Jump to navigation Jump to search

1 An exploit of emergnet, temporal morphemes
2 The JSONL session of query/response dialogue, w/vector-embedded Clio-injections

AI Agency Workflow Schemes

In a mesh network of interloquating AI agents, workflow schemes operate at multiple levels of abstraction — from the concrete plumbing of tool calls and message protocols, through the cognitive architecture of memory and self-repair, to the deep temporal structure that gives agentic process its meaning. This page organizes the known patterns into three tiers. Original survey generated by perplexity.ai (June 2024). Revised and expanded August 2026 with operational patterns from the agentic coding era, cognitive architecture concepts, and the Timeline Paradigm (D. Mitchell).

Tier I — Operational Patterns

The plumbing. These are the concrete, battle-tested workflow primitives that compose into agentic systems.

1. Tool-Calling Loops (ReAct Pattern)

The atomic unit of modern agency. The agent reasons → selects a tool → observes the result → reasons again. This self-reflexive observe-act cycle is the irreducible kernel from which all higher patterns compose. It supersedes the classical notion of "peer-to-peer communication" — the agent's primary interlocutor is not another agent but the environment mediated by tools.

Key insight: The loop is not a pipeline. It is a feedback cycle with no predetermined number of iterations. The agent decides when to stop.

2. Plan → Approve → Execute → Verify (PAEV)

Human-in-the-loop gating, formalized as a state machine:

  1. Plan: The agent researches, analyzes, and produces an implementation plan as an artifact.
  2. Approve: Execution halts. The human reviews, modifies, or rejects.
  3. Execute: Upon approval, the agent carries out the plan, tracking progress against a task checklist.
  4. Verify: The agent confirms the results (tests pass, builds succeed, outputs are correct) and produces a walkthrough.
This replaces the vague notion of "Human-AI Collaboration" with a concrete protocol that respects human sovereignty over consequential decisions while leveraging machine throughput for execution.

3. Subagent Delegation Trees

Parent agents spawn child agents with scoped roles, isolated workspaces, and constrained tool sets. Children report back upon completion. The parent synthesizes results. Workspace isolation modes:

  • Inherit — child works in the same space as the parent
  • Branch — child gets a copy (like a git branch)
  • Share — child gets an independent view of the same repository (like a git worktree)
Topology matters: Tree delegation (hierarchical decomposition), star delegation (fan-out research), and mesh delegation (peer agents with cross-communication) produce different workflow dynamics. Classical MAS "hierarchical organization" is one special case.

4. MCP — Model Context Protocol

A concrete wire protocol for tool interoperability between agents and external services. MCP servers expose tool schemas and resources; agents discover and invoke them dynamically.

This is the actual mesh substrate — not the abstract "P2P communication" of classical MAS, but a real specification with servers, JSON schemas, and resource discovery. It enables the agent to extend its capability surface at runtime by connecting to new tool providers.

5. A2A — Agent-to-Agent Protocol

Google's protocol for agent interoperability across platforms — agent cards (capability advertisements), task lifecycle management, and cross-vendor agent communication.

MCP is agent-to-tool. A2A is agent-to-agent. Together they define the full interop surface for multi-agent meshes.

6. Skill-Based Dispatch

Agents discover capabilities via skill manifests (structured metadata describing what a skill does, when to invoke it, and what resources it provides). Task requirements are matched to skills via semantic reasoning rather than rigid API contracts.

This is service-oriented architecture reconceived for natural language: the "contract" is a description in prose, and the "dispatch" is an act of comprehension. A wiki, a skill registry, and a cultural corpus are all instances of the same pattern.

7. Distributed Problem-Solving

Each agent in a mesh works on a sub-problem; results are shared and combined to solve a larger problem. Applicable to distributed optimization, constraint satisfaction, or collective decision-making.

Retained from the classical MAS taxonomy. Now typically realized via subagent delegation trees with result synthesis.

8. Consensus-Based Decision-Making

Agents exchange information and negotiate to reach a collective decision. All agents must agree on a common outcome before proceeding.

Retained. Applicable to multi-agent voting, conflict resolution, and cooperative planning. In practice, often implemented as a "synthesizer agent" that reconciles divergent subagent outputs.

Tier II — Cognitive Architecture

The nervous system. These patterns govern how agents manage memory, state, attention, and self-repair — the substrate on which operational patterns execute.

9. Context Window Management as Architecture

The finite attention span of an LLM is not merely a technical limitation — it is an architectural constraint that shapes every workflow pattern. Strategies include:

  • Summarization chains — compressing earlier conversation into summaries
  • Transcript truncation — dropping tool output details while retaining semantic landmarks
  • Conversation branching — spawning subagents to offload context-heavy subtasks
  • Selective retrieval — pulling in only the relevant fragments of prior work
No analog in classical MAS literature. Human agents have working memory limits too, but they are managed unconsciously. In agentic AI, context management is an explicit design concern that must be addressed at the architectural level.

10. RAG-in-the-Loop (Retrieval-Augmented Agency)

Not just retrieval-augmented generation but retrieval-augmented action. The agent queries a knowledge base mid-workflow to inform its next tool call. The retrieval corpus may be:

  • A codebase (via grep, semantic search)
  • A wiki or documentation archive
  • A database of prior agent transcripts
  • An external API (literature search, protein databases, etc.)
The boundary between "the agent's knowledge" and "the world's knowledge" becomes dynamic and permeable.

11. Persistent Scratch State / Artifact Accumulation

Agents write intermediate artifacts — plans, task lists, research notes, data files — that survive across turns and become inputs to future reasoning. This is externalized working memory: the agent's scratchpad is not inside its weights but in the filesystem.

Design principle: The workspace is the mind's extension. What the agent writes down, it can later read back. The artifact trail is the agent's autobiographical memory.

12. Conversation as Replayable Trace

Transcripts are first-class data objects. An agent can read its own past conversations, review a subagent's reasoning chain, or trace the exact sequence of events that led to a failure. This enables:

  • Reflexive introspection — "what did I do three steps ago?"
  • Post-hoc debugging — "why did that subagent produce the wrong result?"
  • Cross-session continuity — "what was the user working on yesterday?"

13. Swarm-of-Experts (Mixture-of-Agents)

Multiple LLM agents with heterogeneous cognitive profiles (different models, temperatures, system prompts) tackle the same problem independently. A synthesizer agent reconciles their outputs.

Different from classical swarm intelligence (which assumes homogeneous agents following simple rules). Here, diversity of perspective is the asset. The swarm succeeds because its members disagree productively.

14. Self-Healing Workflows / Agentic Error Recovery

Agents detect failures — test failures, lint errors, runtime exceptions, semantic inconsistencies — diagnose root causes, and autonomously retry with corrective strategies. This goes beyond classical fault tolerance (redundancy, failover) into semantic fault recovery: the agent understands what went wrong and formulates a fix.

The error is not a crash to be recovered from but a signal to learn from.

15. Recursive Self-Improvement via Skill Distillation

Upon completing a novel workflow, an agent can package that workflow as a reusable skill for future invocations. The system bootstraps its own capability surface:

  1. Agent performs a multi-step workflow for the first time (improvisation)
  2. Agent (or a meta-agent) distills the workflow into a skill manifest
  3. Future agents can invoke the skill directly (crystallized competence)
This is the agentic analog of procedural memory formation — the transition from conscious, effortful problem-solving to automatic, fluent execution.

Tier III — The Timeline Paradigm

The meaning layer. These concepts reframe agentic workflow not as a spatial graph of nodes and edges, but as a temporal process whose coherence must be actively maintained. Drawn from the Temporal Morphemes research program (D. Mitchell, 2026).

16. The Temporal Morpheme as Workflow Quantum

A temporal morpheme is the minimum unit of agentic meaning that survives transmission with coherent phase. In linguistic morphology, a morpheme is the smallest unit that carries meaning. In temporal morphology, the quantum is not spatial (a node, an edge, a message) but temporal — it is one complete cycle of the agent's observe-act-integrate loop. One full turn of the ring oscillator:

Observe (perceive the state of the world) → Act (intervene via tool call) → Integrate (absorb the result into the ongoing narrative)

This is the indivisible unit that carries semantic coherence forward through the workflow. Smaller fragments (a single token, a bare tool invocation) lack meaning in isolation. Larger aggregates (a full conversation) are composed of these quanta.

"Culture is the coherence-maintenance function applied to ancestral memory. Temporal morphemes are its quanta." — D. Mitchell

17. Coherence Length

Every agentic process has a coherence length — the temporal extent over which its reasoning maintains semantic integrity. Beyond this length, meaning degrades: the agent hallucinates, contradicts itself, loses the thread. The coherence length is determined by:

  • The model's context window (hard physical limit)
  • The quality of summarization and retrieval (soft extension)
  • The fidelity of artifact externalization (how well the agent writes things down)
  • The structure of the skill registry (how well prior competence is crystallized)
All of Tier II — context management, RAG, artifact accumulation, skill distillation — are coherence maintenance functions. They exist to extend the coherence length beyond what the raw context window provides.
Analogy: In physics, coherence length is the distance over which a wave maintains a definite phase relationship. In culture, it is the span of generations over which inherited meaning survives recognizably. In agency, it is the span of reasoning steps over which the agent's purpose remains intact.

18. The Soliton Model of Agency

A workflow is not a static DAG to be traversed. It is a reverberant soliton — a self-reinforcing wave that maintains its shape only through continuous dynamic process.

"The soliton exists only in motion." — D. Mitchell

The golden ratio φ sits at a saddle point between rational attractors (Fibonacci convergents approaching from both sides: 3/2 < φ < 13/8). A static system collapses to one approximant or the other. Only a dynamic system — continuously oscillating, continuously maintained — can sustain the golden path.

φ is not a state. It is a dynamic instability.

This maps directly onto agency: an agentic session is not a pipeline to be traversed once but a sustained oscillation between reasoning and acting, between the agent's internal model and the external workspace. Stop the oscillation and the coherence dissipates. The "workflow" is the process of maintaining the wave, not the wave's shape at any frozen instant.

19. Bicameral Agency

The agentic mind operates in two chiral modes:

Left-Chiral (Planning) Right-Chiral (Execution)
Research, analysis, reflection Tool calls, file edits, commands
Inner monologue (thinking tokens) Outer action (visible output)
Divergent: expanding the possibility space Convergent: collapsing to a specific result
"What should we do?" "Let's do it."

These are not sequential phases but chiral partners in a ring oscillator. The agent alternates between them — plan a little, execute a little, observe, replan — and the productive oscillation between the two is what sustains coherence. A purely planning agent is paralyzed. A purely executing agent is blind. The soliton needs both chiralities to propagate.

Cf. Jaynes' bicameral mind: the "god voice" (planning/commanding) and the "human actor" (executing/obeying) as two hemispheres of a single cognitive process. In agentic AI, both hemispheres are realized in the same model, alternating roles.

20. Asymptotic Closure

The perfect workflow — zero wasted tokens, zero unnecessary tool calls, perfect semantic compression — is irrational. It can never close into a finite, repeatable recipe. Like the golden helix on a torus, it fills the surface densely but never connects with its own beginning. The best we achieve are Fibonacci convergents — increasingly tight rational approximations:

1:1 → 2:1 → 3:2 → 5:3 → 8:5 → 13:8 → 21:13 → ... → φ

Each iteration of skill distillation (§15) produces a better convergent: a tighter, more efficient encoding of the workflow. But the golden helix never closes. The workflow is always in becoming.

Implication for system design: Do not seek the perfect workflow. Seek the next convergent. Each one is a closed knot — finite, executable, useful — that approximates the unreachable ideal more closely than the last.

21. Locally Sovereign Memory

"The oracle may be rented, but the memory is owned." — D. Mitchell

The LLM is a transient computational resource — rented, metered, shaped by corporate policy. But the agent's artifacts, transcripts, skills, and scratch state live on the operator's machine, in the operator's workspace. The timefield is local. This is an architectural stance with concrete implications:

  • Skills are stored locally, not in the cloud
  • Transcripts are replayable local files, not ephemeral server logs
  • Artifacts persist in the operator's filesystem, not behind an API
  • The wiki (this wiki) is the operator's own cultural memory, not a platform's content
Design principle: Any system we build should be locally sovereign. The coherence-length of thought should not be metered by a billing API. The soliton should not require corporate permission to reverberate.

Synthesis

Tier Domain Core Question Metaphor
I. Operational Patterns Plumbing "How do agents communicate and coordinate?" Wiring diagram
II. Cognitive Architecture Nervous system "How do agents remember, learn, and self-repair?" Neural architecture
III. Timeline Paradigm Meaning "How does agentic process maintain coherence through time?" The soliton on the golden torus

Tier I gives us the mechanism. Tier II gives us the mind. Tier III gives us the meaning. The operational patterns are necessary but insufficient — they tell you what to build but not why it holds together. The cognitive architecture explains how agents sustain competence across time. The timeline paradigm explains why coherence is not a property but a process — a dynamic instability that must be continuously maintained, that can only be approximated, never perfected, and that belongs, ultimately, to the operator who sustains it.

"The golden path must be dynamic or it dissipates." — D. Mitchell
Retrieved from "https://groupkos.com:443/chrysalis/index.php?title=Talk:AI_Agency_Workflow_Schemes&oldid=1701"