Agentic AI

Praxa is a new agent harness that explicitly separates proposal, authority, dispatch, verified effect, and promotion states through deterministic admission and brokered execution. It utilizes external read-back and reconciliation to ensure actions are verified before serving promotion. The system reports evidence from a local audit passing over 1,000 tests and an initial pilot on the Terminal-Bench dataset.

• Explicit state separation prevents conflating agent proposals with verified external effects.
• Deterministic admission and brokered execution enforce strict governance boundaries.
• External read-back and reconciliation steps validate actions before promotion.
• Early evidence includes 1,027 passing unit tests and 89 Workerd tests.
• Initial pilot conducted on Terminal-Bench Core 0.1.1 across 12 curated tasks.

This paper analyzes the distribution of memory usage in long-horizon language agents, revealing that retrieval patterns often follow a heavy-tailed distribution. While core states are retrieved frequently, rare states accumulate in a long tail where prediction errors grow. The study contrasts random-walk policies, which show log-normal artifacts, with semantic LLM policies that exhibit strong truncated power-law core-tail traces, suggesting current evaluation metrics miss critical structural inefficiencies.

• Memory concentration creates a long tail where rare state prediction errors accumulate.
• Semantic LLM policies show stronger heavy-tail traces than random-walk approaches.
• Standard metrics like task success miss structural memory distribution issues.
• Current systems assume uniform memory access, which contradicts observed traces.
• Understanding tail behavior is key to optimizing long-horizon agent reliability.

A new paper introduces FedCausalCompose, a framework that enables modular LLM agents to distinguish correlation from causation during cross-module planning. Standard observational world models fail to identify valid intervention paths, such as whether payment actually authorizes shipment or if inventory mediates the effect. By using local actions as intervention-response evidence, this approach resolves irreducible errors found in traditional observational traces.

• Observational traces fail to distinguish valid causal links from spurious correlations in modular systems.
• FedCausalCompose uses local intervention evidence to map cross-module dependencies accurately.
• Prevents irreducible planning errors when actions in one module gate transitions in another.
• Shifts agent design from passive pattern matching to active causal reasoning.
• Essential for complex workflows like order, payment, and inventory orchestration.

This paper benchmarks off-the-shelf Qwen3 models (0.6-8B) on four agent harness microtasks, including command approval and memory writing, against practitioner-defined thresholds. The study reveals an eligibility gap where even the best-performing FP16 greedy configurations fail to clear confidence bounds anchored to cheap non-LLM baselines. It further investigates whether quantization mitigates these failures or exacerbates them in real-world agent workflows.

• SLMs struggle to meet strict confidence thresholds for agent microtasks without tuning.
• Benchmark uses fixed prompts and auto-metrics anchored to non-LLM baselines.
• Quantization effects on eligibility gaps are explicitly tested in this study.
• 0.6-8B Qwen3 models show significant gaps in command approval and memory tasks.

This paper formalizes budget conservation for AI agents delegating work across failure-prone workers, addressing overspend risks from lost replies, timeouts, and message duplication. It proposes using quantized resource vectors as exclusive escrow credits that move through a delegation DAG. Before dispatch, branches convert credit into operation reservations bound to lineage, epoch, and idempotency keys to ensure precise cost tracking despite network instability.

• Prevents budget overspend in multi-agent systems when replies are lost or duplicated.
• Uses exclusive escrow credits to enforce strict resource allocation via delegation DAGs.
• Binds operations to lineage and idempotency keys for fault-tolerant accounting.
• Addresses edge cases like timeout effects and message repetition in distributed AI.