Skip to content

AI Platform Meeting - 2026-06-09

Info

This is a meeting summary kept for institutional memory. It records a conceptual walkthrough, not a final decision. When a concrete choice is made (for example, a model-agnostic SDK), capture it as an ADR under records/adr.

  • Date: 2026-06-09 (Tuesday), 08:00-11:00
  • Presenter: Özcan Övünç
  • Main participant: Nur Dinçer
  • Also invited: Osman, Osmank
  • Topic: Reference architecture for an enterprise agentic AI platform

Overview

The goal is to design an enterprise-grade agentic AI platform for regulated customers while reusing the existing Appcircle infrastructure as much as possible (database, DevOps, programming languages) instead of rewriting it.

The session walked through the platform as a set of logical layers. The layer boundaries are conceptual: in practice these responsibilities are interleaved and several can live in the same repository. They were separated here to make each responsibility and its purpose clear.

Architecture Layers

Control Plane (Governance)

The central place where all policy enforcement happens. Compliance rules (KVKK, BDDK, BTK) are entered here as policies, evaluated with OPA and Rego. Example policies: masking a credit card so only the last four digits are visible, or preventing sensitive locations from ever being written to memory. For regulated customers this layer is required from day one.

Evidence Fabric

A hashed, immutable audit log of every decision: what was decided, based on which reasoning, who approved it. This exists for audits and regulatory inspection. Like the Control Plane, it is a must-have for regulated customers.

Runtime

Agent runtime, comparable to the process management of an operating system. It starts a process for an agent, runs it to completion, and manages the agent lifecycle.

Coordination (Orchestration)

Manages the team of agents and the workflows between them. Workflows are modeled as a deterministic state machine, and each agent's status within a workflow is tracked as a finite state. Pre-built workflows can be prepared in advance.

Model Plane

Model adapters. The platform is model-agnostic but opinionated. It supports a mixture of experts and routes requests to a model based on a confidence score (simple tasks to a cheaper model, reasoning-heavy tasks to a stronger model such as Opus).

Memory

The layer that needs the most engineering. It follows a human-memory metaphor with four types:

  • Working memory - short-lived, tied to the current task.
  • Episodic memory - past events and customer-specific history; can be modeled per organization, sector, or customer.
  • Semantic memory - learned facts, procedures, and corporate rules (do/don't).
  • Procedural memory - step-by-step governance ("first do this, then this").

For retrieval, a hybrid mechanism is recommended. In internal experiments, an LLM-friendly wiki plus Markdown/SML files produced better results than vector-based RAG, and were easier to maintain and parse. Memory can be structured hierarchically: product level, then sector (banking, insurance, telecom), then individual company.

Integration Layer

Adapters for protocols such as MCP and agent-to-agent communication. The standards here are maturing, so this is a relatively easy layer.

Security (Cross-Cutting)

AI identity, observability, and PII handling run across all layers. NLP/NER models detect and anonymize sensitive data, but they struggle with out-of-vocabulary terms (for example, a brand name used with no surrounding context), so an additional LLM-based check may be needed.

Key Concepts

  • Brain/Hands separation (per Anthropic guidance): do not give execution rights to a "brain" agent, and do not give reasoning rights to an "executor" agent.
  • Team of agents: a minimum of about 5 agents, with around 8 as the ideal. Smaller, well-scoped agents increase determinism and accuracy; overloading a single agent with many roles raises hallucination.
  • Agent loop: Perceive → Reason → Act → Reflect → Critic.
  • Reflection: updates memory, but not instantly. Updates are batched (for example, a nightly job using a random forest over the day's queries) before being written back.
  • Critic: persona-based verification with a veto/retry mechanism. Retry count is configurable; once a confidence score drops below a threshold the flow exits the loop and goes to human approval, which avoids an infinite retry loop.
  • Context replay: past decisions can be replayed step by step. Keeping a memory snapshot for an exact replay is costly, which is a known trade-off.

Direction and Decisions

  • Deliver the platform to customers as an SDK (.NET, Python, and Node SDKs already exist), configured during onboarding. The SDK is roughly 90% ready.
  • The infrastructure stays the same across customers; what differs from one company to another is mainly the skill/agent set.
  • There are few ready-made enterprise options: AWS Bedrock and a Microsoft agentic/policy product. Even so, the Memory, Coordination, and Control Plane layers require hands-on engineering; Integration and Model Plane are comparatively easy.
  • For language and tooling: stay model-agnostic, keep the deterministic workflow/state-machine approach, and use scripting languages (such as Python) where they fit.

Open Questions and Next Steps

  • Strengthening PII detection beyond NER models (whether an additional LLM pass is needed).
  • The cost trade-off of memory snapshots for context replay.
  • Detailed design of inter-agent communication (model plane, runtime invocation).
  • Next: Nur to digest the concept, after which the observability/management side and the lower-level details will be discussed.