research: ACG formalism adoption for engine architecture vocabulary and agent pruning #848
Description
Summary
Evaluate and adopt the Agentic Computation Graph (ACG) formalism from the IBM/RPI workflow optimization survey as the formal vocabulary for SynthOrg's engine architecture.
Research Source
- From Static Templates to Dynamic Runtime Graphs (arXiv:2603.22386, IBM Research + RPI)
- Companion repo: github.com/IBM/awesome-agentic-workflow-optimization
Research Scope
1. ACG Formalism Mapping
| ACG Concept | SynthOrg Equivalent |
|---|---|
| ACG Template | Company YAML config (reusable executable specification) |
| Realized Graph | Runtime execution state (instantiated for one run) |
| Execution Trace | Observability events + costs (states, actions, observations) |
| Nodes | Atomic actions: LLM calls, tool use, validation |
| Edges | Task dependencies |
| Scheduling Policies | Loop selector, routing strategies |
2. Structural Credit Assignment
Open problem -- determining which agent or workflow component caused a failure:
- Per-agent contribution scoring in multi-agent task completion
- Error attribution when tasks fail in a pipeline
- Integration with existing observability event system
3. Agent Pruning / Dropout
Evaluate AgentDropout and Adaptive Graph Pruning techniques:
- Dynamic team sizing based on task complexity
- Removing redundant agents mid-execution
- Integration with HR module (performance tracking informs pruning decisions)
4. Key Survey Findings to Validate
- Structural improvements yield greater gains than prompt refinement when scaffold is poorly matched
- Strong verifiers enable more aggressive graph mutation
- Selection/pruning from a super-graph beats unconstrained generation
- Quality-cost tradeoffs must be explicit with hard budget caps
Deliverables
- Architecture document mapping ACG vocabulary to SynthOrg modules
- Recommendations for structural credit assignment integration
- Evaluation of agent pruning applicability to HR module
Related Issues
- research: audit engine against multi-agent failure patterns (swarm drift, microservices anti-patterns) #690 (audit engine against multi-agent failure patterns)
- feat: coordination metrics from distributed systems theory (Amdahl ceiling, straggler gap) #703 (coordination metrics from distributed systems theory)
- research: five-pillar evaluation framework for HR performance tracking #699 (five-pillar evaluation framework)