Agent teams assembled by configuration that review their own work

The task

Ad-hoc LLM chains do not scale into a product platform: no task routing, no review of agent output, nothing accumulated between runs, and autonomy is all-or-nothing. We needed one core for several products — engineering, research and cognitive agent teams on a shared platform with quality control and managed autonomy.

What we built

• A declarative platform: agents, teams and pipelines are defined in YAML, and the execution graph is built at runtime. A new agent team is added by configuration, not by rewriting orchestration code.
• Quality is held by 10 agent-critic pairs, from tech lead to security red team. Critics score outputs across several dimensions and drive iterative rework; unresolved disputes go to the orchestrator.
• An agent registry matches tasks to executors by capability, role and tags; only knowledge that passes validation is committed to shared memory.
• Four memory circuits: episodic (Redis), semantic search (Qdrant), knowledge graph (Neo4j), tasks and audit log (PostgreSQL).
• Five autonomy levels with self-adjustment and human escalation; a meta-agent tracks team metrics and suggests topology changes. Managed via 34 MCP tools and a monitoring dashboard.

Outcome

25,000+

lines of Python in the platform core across 134 modules, counted from the repository

29

declarative agents in three ready-made teams — engineering, cognitive, research; counted from YAML specs

223

automated tests at three levels: unit, integration, end-to-end; counted from the test suite

2.5 weeks

from first commit to a working core with memory, agent-critic pairs and a dashboard, per repository history

More cases