Agentic AI Core
A foundation layer for autonomous agents
An LLM planner decomposes the goal, an executor runs tools with validation, timeouts and retries, an evaluator checks the result, and memory carries context forward.
- 01Goal
- 02Planner
- 03Executor
- 04Evaluator
- 05Memory
- 06Result
Research agent-architecture papers and summarise the top 3
- ○search recent papers
web_search - ○fetch the top results
http_get - ○write the summary
file_write
goal met · 3/3 subtasks · summary written
→ Result returned with a full trace.
- Roles
- Planner · Executor · Evaluator · Memory
- Memory
- FAISS · Chroma
- Status
- In design
- Year
- 2026
01 Problem
Most “agents” are a prompt in a loop: no plan, no retries, no memory, and no way to tell whether it worked. Agentic Core is the layer underneath — explicit roles, typed contracts and observability from day one — so domain agents can be built on top of it.
02 Approach
Plan
The planner turns a natural-language goal into a structured plan of subtasks with tool assignments, and replans with the error as context when a step fails.
Execute
Tools are validated with Pydantic and run with timeouts and retries: file read/write/search, sandboxed Python, web search and HTTP. New tools are registered through YAML config.
Remember
Short-term task state plus long-term vector memory (FAISS locally, Chroma for persistence).
Evaluate
A self-evaluation step decides whether the goal was met and triggers adaptation when it wasn’t.
Observe and constrain
Structured logs, traces and checkpoints; a written threat model; LLM providers abstracted so GPT-4 or a local model can sit behind the same interface.
03 Results
- Agent roles
- 4
- Built-in tools specified
- 7
- Design specs incl. threat model
- 8
- Product requirements, system architecture, technical design, API contracts, memory schema, evaluation plan, implementation guide and threat model written.
- Deliberate choices: thin wrappers instead of a heavy framework, explicit state, no swallowed exceptions.
- Next: FastAPI service, Chroma persistence, RAG over documents, multi-model routing.
Status: specification complete, implementation in progress.
04 Stack
- Python
- Pydantic v2
- FAISS
- Chroma
- structlog
- FastAPI
- Swappable LLMs