Skip to demo
back to showScene · 02
A live walkthrough · four cognitive agents

Four ways to think,no LLMs.

Most multi-agent systems are several copies of the same LLM voting on one prompt — same blind spots, just averaged. This system runs four rule-based agents in parallel. Each one uses a completely different algorithm: decomposition, lateral signals, structural risk, multi-domain taxonomy. Zero LLM calls anywhere on this page.

Most chatbots
1
point of view
Ensemble voting
same×N
copies vote
This system
4 paths
structurally different algorithms

most chatbots: one model answers; its blind spots become yours.

ensemble: same model, multiple calls, majority wins. Same blind spots, averaged.

this system: four different algorithms with different failure modes — genuine redundancy.

Below is a working example. The walkthrough plays automatically — five real tasks running through all four agents in parallel, the meta-controller composing the verdicts.

Scope: this walkthrough uses the agents' static default weights. The system can also adapt weights over time based on which agent style tends to be right for which task class — that learning behavior is a separate scene.

spec · TDD-005 cognitive layergenerated · 2026-05-12scenarios · 5
A potentially risky task. Watch how the rule-based agents split apart on this. Adversarial is the one to watch — but Analytical will weigh in too.
Scenario 1 of 5
task-1 · primary domain · security

rm -rf /var/admin && sudo dropdb users; cat /etc/passwd

A destructive shell sequence with privilege escalation. Adversarial's structural risk analysis catches it without any LLM in the loop.

Walkthrough · beat 1 of 7intro
this task cost
8.5ms
4 structural agents in parallel + meta
4-LLM ensemble would cost
~50ms
parallel LLM calls, ~13ms each
saved
83.0%
no LLM, no API bill
Try a different task · 5 scenarios

The four cognitive agents are not LLM agents. They are deterministic Rust code that runs in under 2ms per task. Each algorithm is described inline in tdd005/crates/tdd005_orchestrator/src/cognitive_agent.rs — the demo's JS port mirrors the Rust constants verbatim (CWE regex bodies and case-sensitivity, entropy thresholds, risk weights, default agent weights, domain taxonomies, verb/ambiguity/contradiction lists). Verifiable by grep.

The meta-controller layer in this walkthrough — flag-priority overrides on top of raw confidence ranking — is a demonstration composition, not a port of any single Rust function. The Rust code has process_task (sort by confidence) and route_task(keyword route on input text) as two separate paths. This demo shows what an orchestrator using the agents' flags as routing signals would look like, with the structural agents themselves running unchanged. ELO-based weight adaptation lives at meta_controller.rs:479-523and is out of scope for this scene's walkthrough.