I wanted to explore what changes when a business brief gets examined by several specialist perspectives instead of one model trying to cover everything at once.
← Back to all work
Engineering case study · AI Systems Engineering · 2026
Board of Directors AI
A parallel multi-agent system where eight executive-role agents analyze a business brief concurrently across finance, operations, technology, product, marketing, risk, legal, and HR.
Central idea
What this system investigates.
Multiple perspectives only create value when the system reconciles disagreement instead of concatenating opinions.
Problem
Why this system exists.
Single-model business analysis compresses finance, operations, technology, product, and risk into one undifferentiated response. Parallel agents introduce depth but also introduce state conflicts when concurrent branches share output fields and complete at different times.
System architecture
How the system is organized.
- Read brief
- Parallel reactions
- Assign directives
- Run tiered analysis
- Evaluate and revise
- Reconcile contradictions
Implementation
Methods and engineering choices.
Department agents operate in dependency-aware parallel tiers with explicit synchronization gates that wait for sibling branches before advancing. Per-agent state prevents concurrent branches from overwriting one another. Evaluation loops can request revisions before a branch result is committed. The reconciliation layer actively checks the final synthesis for contradictions.
Results and evidence
What the current record demonstrates.
- 21-commit public repository with parallel-flow tests covering branches that complete at different times
- Per-agent state design prevents concurrent branches from overwriting one another during parallel execution
- Staggered revision tests exercise the evaluation loop against branches with varying completion timing
- API authentication, rate limiting, retries, and output isolation are documented as first-class design decisions
Verification
Checks performed and their scope.
- Parallel-flow tests exercise branches completing at different times and staggered revision behavior.
- Authentication, rate limiting, retries, and output isolation are documented as implementation controls.
Limitations
What this work does not yet establish.
- Output quality against expert review has not yet been benchmarked.
- Revision loops and model-call reduction remain open engineering work.
Engineering decisions
Decisions that define the model boundary.
- Redesigned shared state structure before introducing parallel execution — not after hitting race conditions.
- Used explicit synchronization gates rather than assuming graph fan-in behavior handles ordering correctly.
- Deferred bidirectional revision loops until their added complexity is justified by measured output quality.
A note from the build
What stayed with me.
Parallel opinions introduce their own problems: agents can disagree, clash over shared state or leave contradictions for the final answer to smooth over.
I learned that coordination is part of the system. Clear ownership, synchronization and a real reconciliation pass matter as much as the individual agents.
Next step
What I would investigate next.
Test the quality of final recommendations against expert review and reduce total model calls without weakening cross-functional analysis depth.
Source and related work