← All work2025 · AI Agents · Infrastructure

Declarative Multi-Agent Framework

Agents, teams, and knowledge bases declared in one TOML file — no orchestration code.

RoleAuthor
StackPython 3.11+ · TOML · Agno
Year2025

Client work — case study anonymized; architecture and metrics are real.

Configuration as the interface

Fifteen pluggable tools — from web search and finance data to Python execution and PubMed — three team orchestration modes (route, coordinate, collaborate), eight supported models, and RAG over CSV, PDF, and web sources. All of it declared, not coded, with strict startup validation that fails with clear errors instead of running half-configured.

team.tomltoml
[team]
name = "research-duo"
mode = "coordinate"          # route | coordinate | collaborate

[[agents]]
name  = "researcher"
model = "gemini-2.5-pro"
tools = ["exa", "wikipedia", "pubmed"]

[[agents]]
name  = "analyst"
model = "llama-3.3-70b"
tools = ["pandas", "calculator", "python"]

[knowledge.papers]
kind    = "pdf"
sources = ["./corpus/*.pdf"]
rerank  = "cohere"           # qdrant retrieve -> cohere rerank
A complete two-agent research team, declared

Built with

Every technology this project runs on, and the job it does here. Each one links to the other projects it appears in.

Languages & runtimes

Python 3.11+↗Primary backend language: agent services, ETL pipelines, eval harnesses and CLI tooling. Typed end to end with Pydantic v2 models, checked under mypy --strict.TOML↗Declarative configuration format for agents, teams, tools and knowledge bases, validated strictly at startup.

AI & agent systems

Agno↗Agent runtime under production assistants — wrapped in an in-house resilience, telemetry and auth layer rather than called raw.Google Gemini↗2.5 Pro, Flash and Flash-Lite for reasoning and conversation across most agent products; vision models for reading colorimetric test strips from uncontrolled phone photos; image generation for visual try-on.Together · DeepSeek · Llama 3.3 70B↗Alternate providers behind a pluggable model loader, swappable by configuration alone.

Retrieval, ML & vision

RAG pipeline↗Cache → vector retrieve → rerank, with O(1) partition invalidation so one source can be re-indexed without a full rebuild.Qdrant↗Vector store for partitioned knowledge bases behind retrieval-augmented agents.Cohere rerank↗Second-stage ranking after vector retrieval, ahead of context assembly.Exa↗Live web research inside agent stages — with a stubbed provider so the service still boots and runs with no key present.

By the numbers

15Pluggable tools
3Orchestration modes
8Supported models