Exactly... this probabilistic / deterministic tension is real.
An example of this is this experiment with a governed Credit Risk Workflow executed by Codex and Claude Code.
In one case, Codex resolves the workflow over a filesystem substrate. In another, Claude Code resolves the same workflow materialized in SQLite.
The flow is basically:
Producer -> Executor -> Reviewer
The interesting part is separating when the model acts as the primary resolver, when deterministic code becomes the core, and when code is just an IO/instrumental layer.
https://t.co/UfLHNdw9Tr
We are seeing an explosion of memory, sandbox, and orchestration solutions for AI agents.
Useful, but increasingly fragmented into proprietary operational spaces.
Throughout computing history, many distributed solutions for massive use eventually converged toward protocol layers.
Maybe agent memory, execution context, sandboxing, and workflow orchestration are moving in the same direction.
As LLMs become more capable, and agents become their operational layer, we may need a protocol layer for governed workflows.
- projects as composable workflows
- context as addressable structure
- domain knowledge and rules as reusable artifacts
- complex intent expressed as governed declarative pipelines
- code as a dynamic instrument
- results as traceable outputs
- agents as governed executors
- model-agnostic execution
The future may not be only about better agents.
It may also be about better operational spaces for agents to work in.
Does it make sense to think of agent workflows as an operational space that may need a protocol layer?
#AI #LLM #AIAgents #AgenticAI #ContextEngineering #AIInfrastructure #GovernedWorkflow
I agree with the direction.
But I would go a bit further: create a structure with a clear separation between governance, data, logic, and results.
In this model, the LLM acts as a governed resolver within an operational structure, where parts of the pipeline can be resolved by instrumental code or by inference, according to the governance rules.
A practical example:
https://t.co/e4OaZ8qbP4
@intology I did some experiments with Claude Code and Codex in HTL Optimization Campaign Simulation, with the main objective of testing a protocol structure.Same workflow executed by:
Codex:
https://t.co/NrBBp29MBe
Claude Code:
https://t.co/l3j5PB5Mj1
Skills as a Protocol Layer
Skills are an important evolution in the AI ecosystem.
A skill can act as a protocol instruction layer.
The skill provides the operational grammar.
The workflow provides the domain logic.
This means that business knowledge does not need to become a new specialized skill for every domain, rule, procedure, or operational context.
Instead, it can live inside the workflow itself as an extensible logic layer.
In this structure:
- Skill as protocol instruction.
- Workflow as governed execution.
- Operational space as addressable structure.
- Logic as addressable knowledge.
- Model as governed resolver.
- Agent as executor.
- Substrate-agnostic persistence.
This is the direction I have been exploring with AIURM/AIUAR:
A Protocol Layer for Governed Cognitive Workflows.
#AI #LLM #AIAgents #ContextEngineering #GovernedWorkflow
HR Analysis Workflow – (AIURM/AIUAR + Claude Code)
Substrate-agnostic practical experiments:
Agent 1: executes a 14-step from the filesystem
Agent 2: materializes the project into a portable JSON file and a portable Markdown file
Agent 3: executes the full pipeline in the JSON file
Agent 4: executes the full pipeline in the Markdown file
step-by-step:
https://t.co/40KTS1oshx
example projects:
https://t.co/Hr6Abo4Ue6
#AI #LLM #AIAgents #GovernedWorkflow #EnterpriseAI
What is AIURM/AIUAR?
Is this a framework?
A DSL?
A specification?
A marker system?
A skill layer?
Prompt engineering?
Context engineering?
A blackboard architecture?
Shared memory?
Agent orchestration?
Maybe one challenge with AIURM/AIUAR is that it is hard to categorize, because it converges, in some dimension, with many already familiar categories.
In that sense, AIURM/AIUAR can be framed as a protocol layer for governed cognitive workflows over an addressable substrate.
https://t.co/SdTDGHqqOd
#AI #LLM #AIURM #AIUAR #AIAgents #GovernedWorkflows #EnterpriseAI
I agree that a governance layer is very important. And I'm intrigued by how little it's discussed.
This is also the direction I have been exploring with AIURM/AIUAR: a protocol-based abstraction layer for complex cognitive workflows, agnostic to domain, substrate, and model.
https://t.co/e4OaZ8qbP4
Paradigm shift:
From frameworks as the runtime
to LLMs as the runtime.
From LLMs as simple calls
to protocol as the execution contract.
From frameworks deciding each step
to protocols expressing the pipeline.
From LLMs executing isolated steps
to LLMs resolving governed workflows.
From complexity embedded in the orchestrator
to complexity governed and resolved at execution time.
What do you think?
Claude Code Resolving a Governed Credit Risk Workflow over SQLite
A 3-agent loop:
Producer → Executor → Reviewer
Three distinct projects operating over the same addressable substrate.
The producer creates pending sessions.
The executor claims and resolves eligible sessions.
The reviewer analyzes completed executions and writes new artifacts back.
Governance modulates execution behavior.
Inferential resolution → model as primary resolver
Deterministic execution → code as core
Instrumental execution → code as IO_ONLY layer
This is the pattern I am exploring with AIURM/AIUAR:
the LLM as a governed orchestrator and executor over an addressable substrate.
No framework.
No orchestrator.
Protocol + Governance + Substrate.
Full demo:
https://t.co/8NXMKoy56f
#AI #LLM #AIAgents #GovernedWorkflow #EnterpriseAI
Structuring AI Workflows with Governance by Design
This experiment uses a complex Small Molecule HTL Optimization Campaign Simulation to demonstrate AIURM/AIUAR as a protocol layer for governed cognitive workflows.
The goal is to show how an executor agent can resolve a multi-step workflow over a persistent, addressable substrate, generating intermediate artifacts, ranked outputs, and operational traces.
Claude Code acts as the executor.
The filesystem acts as the structured substrate.
Governance defines the workflow.
Data, logic, and results become addressable artifacts.
Code is dynamic and instrumental.
All chemistry data, campaign parameters, and outputs are synthetic and used strictly for demonstration purposes.
https://t.co/oLaPsZTrvG
#AI #LLM #AIURM #AIUAR #AIAgents #EnterpriseAI
Experiment scale:
24 JSON files
84K records in one JSON artifact
3.9M lines in one JSON artifact
574 MB of total JSON output
This matters because governed AI execution needs more than a final answer.
It needs persistent state, intermediate artifacts, traceability, and reuse.
Beyond Agent Memory
I’ve been seeing many interesting proposals around protocols for agent memory.
That matters.
But I think we can go one step further: standardizing the agent’s operational space.
Not only what the agent remembers, but where the agent works.
Data, logic, rules, governance, execution state, results, audits, logs, and code artifacts should be persistent, addressable, governable, and substrate-agnostic.
Memory is part of the operational space.
The operational space is the agent’s execution environment.
Here is a practical experiment:
https://t.co/e4OaZ8qbP4
#AI #LLM #AIAgents #ClaudeCode #Codex #EnterpriseAI
Do we need more hardcoded agents?
Or more governed workflows?
Maybe the better question is not:
“Which agent do we need to create for this workflow?”
But:
“How do we structure this workflow so that a general-purpose agent can resolve it?”
#AI#LLM#AIAgents#ClaudeCode#Codex #EnterpriseAI
This also changes the operating model:
From:
- many domain-specific skills
- agent-centric workflows
- monolithic context passed to the model
- persistence in a specific format
To:
- a base set of protocol skills
- project-governed execution
- explicit separation of Data, Logic, and Result as addressable artifacts
- persistence in an addressable, substrate-agnostic operational space
Do we need more hardcoded agents?
Or more governed workflows?
Maybe the better question is not:
“Which agent do we need to create for this workflow?”
But:
“How do we structure this workflow so that a general-purpose agent can resolve it?”
Read more:
https://t.co/e4OaZ8qbP4
#AI #LLM #AIAgents #ContextEngineering #ProtocolEngineering #GovernedWorkflow
One way to frame the shift:
From:
- LLM as a node in a workflow
- logic embedded in code
- fixed orchestration code
To:
- LLM as a governed runtime resolving the workflow
- logic as artifact
- code as an instrument, generated or used dynamically