Agentic Workflows (operational patterns)

Status: Active
Last Updated: April 22, 2026
Tags: agentic, orchestration, governance

---

Summary

This entry describes how agents collaborate, who makes decisions, and which guardrails keep them aligned with GitOps, Ticketing, and Documentation pillars.

Context

Agents operate under TELOS and must learn from every session. These patterns become playbooks that future agents reference when coordinating multi-step infrastructure changes.

---

Observed Workflow Patterns (Active Sessions)

1. "Sequential Continue" Pattern

The user issues a single continue prompt repeatedly. The agent:

1. Reads kb/tasks.md to determine the next unchecked task.
2. Implements it end-to-end (backend + frontend + KB update).
3. Runs get_errors to validate before marking complete.
4. Updates kb/tasks.md session log and task checkboxes.
5. Signals readiness for the next continue.

Key rule: Each continue must result in a shippable, error-free unit of work. Never leave the codebase in a broken state between turns.

---

2. Single-File Project Safety (src/main.tsx > 1000 lines)

• Before inserting any type, function, or component: grep_search for the symbol name to confirm it doesn't already exist.
• Ordering: imports → context/auth → shared types → utilities → page components → auth pages → route definitions → router bootstrap.
• Duplicate risk: After context summarization the agent loses memory of what was written. Duplicate declarations surface as TypeScript errors. Fix by removing the duplicate block, not both.

---

3. Backend-in-Vite-Config Pattern

All API routes registered as Vite middleware in vite.config.ts. Viable in dev/preview; migrate to standalone server before production load.

• Route handler naming: handle<Domain><Action> (e.g., handleAuthLogin)
• Single prisma client at module scope
• Secrets from process.env only — never hardcoded

---

4. KB-First Session Discipline

Every session that adds a feature must:

1. Mark kb/tasks.md checkboxes and add a dated ## Session Change Log entry.
2. Update kb/lessons-learned.md with new patterns or gotchas.
3. Update kb/implementations.md if an architectural decision was made.
4. Update kb/agent.md if environment/routes changed significantly.

---

Core Orchestration Patterns

1. Orchestrator–worker separation – A single coordinator agent breaks tickets into tasks, dispatches specialist workers (network, infra, docs), and merges results so ownership stays clear.
2. Evaluator–optimizer loop – Evaluation agents watch GitOps pipelines to catch regressions before commits land in the main branch.
3. Routing/classifier agents – Input classification routes work to domain experts, reducing context switching and ensuring the right skills respond to each issue.
4. Autopilot-assisted design – Borrowing UiPath Maestro, combine rapid iteration with human review so even experimental flows stay explainable.
5. Scaling with Agent 365 principles – Prepare for thousands of agents by enforcing lifecycle states, audit logs, and explainability requirements.

---

Guardrails & Governance

• Transparency gaps, indeterminism, and blurred data-action boundaries are top risks; use low-code workflow definitions to keep the intent + data separation visible.
• Log every agent’s reasoning, input, and output sequence inside Forgejo Issues so “Ticketing is Truth” stays intact.
• Keep Dotenvx-managed secrets scoped and serve them through read-only boundaries; document each boundary inside the KB.

Operational Checklist

• Align every agent with GitOps manifests and validate in staging before promotion.
• Document the agent’s “why” in /kb/ and associate it with the closing ticket.
• Run periodic reviews to spot drift and retire or refactor underperforming agents.
• Before inserting into a large single-file project, grep_search the target symbol name first.
• After every feature implementation, run get_errors before declaring done.
• Every session: update kb/tasks.md + kb/lessons-learned.md + kb/implementations.md.

Next Steps

• Author a diagram that maps these patterns to the actual services (Forgejo, Prisma, TanStack dashboards).
• Link the patterns to the agent dashboard when it ships so each feature reflects a documented workflow.
• Expand the checklist with incident-response drills once the ticketing system is online.

Sources & Related

• Agentic AI Foundation playbooks (AAIF, Agents.md, MCP)
• ServiceNow + PwC orchestration platforms (Agent Studio, Agent OS)
• TechRadar insights on enterprise agentic AI risks
• TELOS doctrine (../TELOS.md)
• Route readers to ../research/ai-server-management.md for the latest citations
• ../implementations.md — Auth, newsletter, single-file architecture decisions (April 2026)
• ../lessons-learned.md — Lessons 16–20 (April 2026 session)

Change Log

• April 22, 2026 — Added "Sequential Continue" workflow pattern and single-file project workflow from active CRM/auth build session.
• January 30, 2026 03:05 AM — Restructured to the KB store format and captured governance-focused workflow patterns.