Overview
The idea is practical: reuse past successful agent transitions to cut redundant turns; experiments on SRDD show token and quality gains but only on similar tasks and a single dataset.
Citations0
Evidence Strength0.60
Confidence0.62
Risk Signals9
Trust Signals
Findings with numeric evidence: 3/3
Findings with evidence refs: 3/3
Results with explicit delta: 5/5
Reproducibility
Status: Partial assets available
Open source: Unknown
At A Glance
Cost impact: 70%
Production readiness: 40%
Novelty: 60%
Why It Matters For Business
Co-Saving can cut token bills and developer compute costs by reusing prior multi-agent transitions, while keeping or improving code quality on similar tasks, so teams can scale automated software generation under a fixed budget.
Who Should Care
Summary TLDR
Co-Saving adds a small memory of past successful agent interactions (called "shortcuts") to multi-agent software-development systems. It ranks shortcuts by value vs cost (time and token usage), applies a dynamic emergency factor tied to remaining budget, and forces termination when interaction cost hits reference limits. On the SRDD software tasks, Co-Saving reports a large cut in token use and higher overall code quality versus prior multi-agent systems, while ablations show shortcut selection and the emergency factor materially affect success and budget completion.
Problem Statement
Multi-agent systems for software development produce good results but often waste tokens and time through redundant interactions. The paper aims to make multi-agent collaboration resource-aware so agents can reuse prior successful transitions to save tokens/time while keeping or improving code quality.
Main Contribution
Introduce "shortcuts": instruction fragments mined from historical multi-agent trajectories that connect non-adjacent solution states and can bypass redundant reasoning steps.
Design a value-vs-cost scoring and filtering pipeline (time, tokens normalized, harmonic mean) plus an "emergency factor" that weights cost more as budget depletes.
Key Findings
Co-Saving reduces token usage versus ChatDev.
Co-Saving improves measured overall code quality versus ChatDev.
Results
| Metric | Value | Baseline | Delta | Split / Dataset | Evidence | Evidence Ref |
|---|---|---|---|---|---|---|
| Token usage reduction vs ChatDev | 50.85% reduction | ChatDev | -50.85% | SRDD (experiments) | Abstract claim: average reduction of 50.85% in token usage | Abstract |
| Overall code quality improvement vs ChatDev | 10.06% improvement | ChatDev | +10.06% | SRDD (experiments) | Abstract claim: improves the overall code quality by 10.06% | Abstract |
What To Try In 7 Days
Log agent interactions as (state, instruction, next state) triples and build a small shortcut index from past successful tasks.
Implement a cheap embedding retrieval (text-embedding-ada-002 or similar) to find reference tasks for new requirements.
Add simple cost filters: estimate token/time cost for candidate shortcuts and drop those exceeding remaining budget; test forced termination thresholds.
Agent Features
Memory
Planning
Tool Use
Frameworks
Is Agentic
Yes
Architectures
Collaboration
Optimization Features
Token Efficiency
System Optimization
Inference Optimization
Reproducibility
Data URLs
Risks & Boundaries
Limitations
Relies on finding similar historical tasks; cold-start tasks get no shortcut benefit.
Embedding-based similarity may miss fine-grained code semantics and produce imperfect matches.
When Not To Use
For novel tasks without historical analogs in the shortcut store.
When budgets are so large that extra reasoning improves quality and cost is irrelevant.
Failure Modes
Applying an incorrect shortcut that produces semantically wrong code despite compiling.
Over-pruning useful interactions and returning incomplete implementations.