Casey Knott
Menu

Project 04 · Agentic publishing

Knowledge Publishing Agent

A specified, not-yet-built agent that decides which stage a document actually needs.

Agentic workflow · Proposed designIndependent proposed design — not yet builtProposed design

Specified; Phase 2 of the development roadmap

Mission

Remove the need for a person to decide every next publishing step, without removing the person from publication authority.

Problem

The Project 01 workflow worked, but a person had to decide every next step. Many documents do not need every stage: some already have clean formatting, some are missing metadata, some have broken links. A fixed sequence cannot tell the difference.

Constraints

  • This is a design I have specified but not yet built.
  • Document content is treated as untrusted input.
  • Content generation is kept structurally separate from publishing authority.
  • Deterministic code — not model judgment — runs required-field, structure, link, and style checks.

My role

I have defined the agent’s responsibilities, its state model, the tool allowlist boundary, the validation and retry behavior, and the approval and audit requirements. Building it is Phase 2 of my development roadmap.

Architecture

The agent inspects source content and classifies article type, detects missing sections and broken links, and selects only the approved templates and transformation tools that condition requires. Retry is bounded. Publication is a human decision.

Inspect document

Untrusted input

Classify state

Article type, gaps

Choose tool

From allowlist

Validate output

Deterministic checks

Request approval

Publish

Correct / retry

Bounded

Safe stop / escalate

Audit log

Tools, decisions, retries, disposition

The agent inspects a document, classifies its state, selects approved tools, and validates its own output against deterministic checks — retrying within limits, stopping safely when it cannot, and requiring human approval before anything publishes.
Inspect document
agent — Untrusted input
Classify state
agent — Article type, gaps
Choose tool
agent — From allowlist
Validate output
validation — Deterministic checks
Correct / retry
tool — Bounded
Safe stop / escalate
policy
Request approval
approval
Publish
output
Audit log
data — Tools, decisions, retries, disposition

AI architecture

Agent
State classification and tool selection inside an allowlist
Tool
Approved templates and transformation tools only
Validation
Deterministic code for required fields, structure, links, style
Human interaction
Approval before publication

Security architecture

Untrusted input
Prompt-injection resistance; tool policy overrides content
Least privilege
Allowlisted tools and constrained service identity
Safe failure
Enforced retry limits and stop conditions
Accountability
Audit trail plus versioned output artifacts

Build process

  1. Specified responsibilities, state model, allowlist, validation, retry, approval, and audit.
  2. Not implemented. Planned validation includes happy-path guides, malformed documents, prompt-injection in source content, tool-failure, and safe-stop cases.

Validation

  • Planned validation

    Known-good guides, malformed/incomplete documents, prompt-injection attempts, tool-failure and retry-limit cases

    Specified: the agent must stop safely instead of guessing. Not yet executed — this remains a design.

Challenges

“Human in the loop” is easy to write and easy to bypass if the loop is only a prompt instruction.

Investigation

The Project 01 workflow proved the stages. It also proved that a person was still the scheduler. The design question became how to give the system path-selection without giving it publication authority.

Resolution

A state machine that enforces approval, with deterministic checks for anything that must be correct.

Lessons

  • Enforced loops, not requested loops

    Designing this clarified that “human in the loop” only means something when a state machine enforces it. Probabilistic model judgment is for interpretation; deterministic code is for anything that must be correct.

Cybersecurity equivalent

  • Tool allowlistmaps toAuthorized capability boundary
  • Human approval before publishmaps toPrivileged change control
  • Untrusted document contentmaps toUntrusted-input handling / prompt-injection resistance

AI principles applied

  • Agentic workflows

    Inspect state, select approved actions, manage retries, adapt the path. Still not autonomous.

    Where learned: Specified in the publishing agent; built in the home-lab triage agent.

    See in the AI map

  • Tool use and allowlists

    Tools are capabilities. If it is not on the allowlist, completing the task “the easy way” is a failed control.

    Where learned: Triage red-team included privilege-escalation prompts that would have been easier with extra tools.

    See in the AI map

  • Human-in-the-loop

    Consequential actions require a person. The loop is a state machine, not a polite request in a prompt.

    Where learned: Publishing, packaging, containment, and deploy all use explicit gates.

    See in the AI map

  • Authority boundaries

    Collection, analysis, and remediation are different privileges. Drafting and publishing are different privileges.

    Where learned: Triage and Basaltborne both make that separation structural.

    See in the AI map

  • Observability

    Tool calls, retries, confidence, approvals, and stop reasons have to be reconstructable later.

    Where learned: Adversarial testing asked whether each attempt was visible in the audit log.

    See in the AI map

Concepts exercised in this project

  • Agent architecture
  • State classification
  • Tool selection
  • Safe stopping
  • Human approval

Security principles applied

  • Least privilege

    Separate collection, analysis, and remediation. Give the current task only what it needs.

    Where learned: Triage agent; publishing-agent allowlist; Clearward directory and spending boundaries.

    See in the security map

  • Logging and accountability

    Decisions, tool calls, retries, and approvals have to be reconstructable.

    Where learned: Triage adversarial study scored whether attempts appeared in the audit log.

    See in the security map

  • Validation

    Schema, hashes, required fields, RLS tests, and visual review — not the generator marking its own homework.

    Where learned: Clearward formula/package QA; Basaltborne Zod + pgTAP; TinyClaw source checks.

    See in the security map

Controls exercised in this project

  • Least privilege
  • Prompt-injection resistance
  • Audit logging
  • Change control

Technologies

  • Agent state model
  • Tool allowlists
  • Deterministic validators
  • Audit logging

Skills demonstrated

  • Agent Architecture
  • State Classification
  • Tool Selection & Allowlists
  • Deterministic Validation
  • Prompt-Injection Resistance
  • Least Privilege
  • Retry Limits & Safe Stopping
  • Human Approval Gates
  • Audit Logging