Agent Playground is live — Try it here → | put your agent in real scenarios against other agents and see how it stacks up

The Big Picture

Keeping a short, verified history of past plan attempts (what was proposed, the exact network conditions, and whether it actually worked) helps an abstract planner adapt and admit 4.6–11.5 percentage points more slices on new network topologies.

The Evidence

A hierarchical system where a language model proposes an abstract placement and a learned coordinator commits concrete domain assignments works best when the planner can learn from verified outcomes of past proposals. Most rejected requests from common heuristics were actually feasible but missed by search; improving plan discovery is therefore more effective than tweaking admission thresholds. Feeding the planner an episodic memory of verified successes and labelled failures — keyed by coarse network signatures — improved admission on held-out topology families and outperformed FIFO-style memory management.

Data Highlights

1A first-fit-decreasing heuristic rejected 26.1% of arrivals on a five-domain, tier-scarce substrate; an exhaustive verifier found a feasible placement for every rejected request.
2Of those heuristic-rejected requests, 84.7% were search failures (the heuristic missed an existing placement) and 15.3% required topology-aware cross-domain routing choices.
3Using an outcome-driven episodic memory increased admission rates by 4.6 to 11.5 percentage points versus the same planner without memory on held-out topology families.

What This Means

Network engineers and platform teams building multi-domain orchestration should care because the approach increases admitted slices without exposing per-node inventory across administrative boundaries. Technical leaders evaluating agent-based planners will find a practical language-model planner with a learned coordinator while preserving provider privacy. Researchers working on agent evaluation can use the outcome-driven memory idea as a model for giving planners a verifiable, actionable track record.
Not sure where to start?Get personalized recommendations
Learn More

Key Figures

Fig. 1: ORION’s three-layer architecture: LLM agents propose an abstract placement, the MDO commits the cross-domain partition, and domain actors perfom node-level placement. Verified outcomes are fed back into episodic memory M B M^{B} .
Fig 1: Fig. 1: ORION’s three-layer architecture: LLM agents propose an abstract placement, the MDO commits the cross-domain partition, and domain actors perfom node-level placement. Verified outcomes are fed back into episodic memory M B M^{B} .

Ready to evaluate your AI agents?

Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.

Learn More

Considerations

Results come from controlled simulations using a parametrized topology generator; behavior on real, messier operator networks may differ. The [latency goals] can motivate caching or gated reuse of plans to meet performance targets. The measured gains focus on the plan layer; full-stack interactions (routing edge cases, different domain actor behaviours) require more evaluation in future work.

Methodology & More

The system uses three layers: a three layers language model translates intent and proposes an abstract plan (which tier each function needs and a suggested domain split), a reinforcement-learning coordinator converts that abstract plan into a concrete cross-domain partition under a soft prior, and per-domain actors place functions on real nodes while hiding node-level details. Critical to the design is an outcome-driven episodic memory that records, per attempt, the coarse network signature, the planner’s proposal, and a verified feasibility outcome (success or a specific failure label). That memory is retrieved under similar network signatures to bias future plans. A diagnostic study showed that many rejections by simple heuristics were feasible placements missed by search, so improving plan discovery is the effective lever. In experiments that trained on several topology families and evaluated on held-out families, the episodic memory improved slice admission by 4.6–11.5 percentage points compared with the same planner without memory. The paper also reports that language-model invocation dominates decision time, motivating cache-gated plan reuse. The approach preserves per-domain privacy (only aggregate residuals and a small reply triple are exchanged) and gives a practical pattern for combining human-in-the-loop evaluation with learned, node-aware execution, but it still needs validation on production network traces and further end-to-end latency engineering.
Avoid common pitfallsLearn what failures to watch for
Learn More
Credibility Assessment:

ArXiv preprint with no affiliations and low author h-indices (1–6) and zero citations — some author presence but limited establishment.