On July 13, three independent research teams released agent reinforcement learning infrastructure within hours of each other. Skyfall AI shipped MORPHEUS, a persistent enterprise simulation benchmark. Stanford published TRACE, a system that diagnoses recurring agent failures and trains targeted adapters. Prime Intellect launched Verifiers v1, a composable evaluation stack that separates tasks, harnesses, and runtimes into interchangeable modules. None of the three teams coordinated. Each solved a different layer of the same problem.

The problem: most agent benchmarks still reset the world after every episode, test a single static task, and bundle their evaluation logic into monolithic codebases. Production agents face none of these conditions. They operate in persistent environments where yesterday’s decisions shape today’s options. They fail for specific, recurring reasons that generic training never addresses. And teams evaluating them rebuild the same infrastructure from scratch for every new benchmark.

MORPHEUS: Environments That Never Reset

Skyfall AI’s MORPHEUS starts from the Big World Hypothesis, formalized by Javed and Sutton in 2024: the world’s complexity exceeds any agent’s representational capacity, so even fixed dynamics look non-stationary to a finite learner. Standard RL benchmarks ignore this by resetting state after every episode. MORPHEUS does not reset.

Each environment is a self-contained TypeScript world plugin that exports Operational Descriptors (ODs), defining step-by-step execution plans for agent capabilities. Non-stationarity comes from two engines running simultaneously. A failure injection engine inserts typed disruptions between OD steps, drawing from eleven failure types (missing data, dependency failures, rate limits, among others) at four preset rates: light (5%), realistic (8%), moderate (15%), and aggressive (30%). An asynchronous configuration shift controller changes failure presets and demand levels at fixed timestamps, running independently of the training loop so agents cannot use update periodicity as a proxy clock.

The evaluation protocol replaces cumulative reward with six metrics: per-configuration reward, adaptation speed, forgetting, recovery time, stability, and performance gap. Adaptation speed, the headline metric, counts steps until a running-average reward reaches half the theoretical upper bound.

The baseline results are the most revealing part, according to MarkTechPost’s coverage. Four algorithm families were tested from a shared supervised fine-tuning checkpoint (Qwen3-14B, initialized from Gemini 3.1 Pro trajectories): PPO, Hindsight Experience Replay (HER), Elastic Weight Consolidation (EWC), and a Latent Context Model (LCM). No single family dominated. On dynamic resource allocation under structured drift, EWC led reward while LCM adapted fastest. On scheduling under drift with delayed effects, HER led reward while LCM lost its edge entirely. PPO and HER adapted only in the first configuration, then flatlined.

Mean performance gaps sat near 1.0 for every method. That number represents the distance between actual performance and the theoretical upper bound. A gap of 1.0 means every algorithm tested is roughly as far from optimal as a random baseline. Current continual RL techniques are not solving the problem MORPHEUS defines. They are exposing how large the gap is.

TRACE: Training What the Agent Actually Lacks

Where MORPHEUS redefines the environment, Stanford’s TRACE redefines the training signal. The core observation, documented in arXiv:2604.05336, is that agentic LLMs fail in patterns. A small set of missing capabilities accounts for most failed trajectories. Direct RL and supervised fine-tuning waste compute because sparse rewards never identify which skill was missing, and broad synthetic data spends budget on skills the model already has.

TRACE runs a four-step automated pipeline. First, contrastive capability analysis splits agent rollouts into successful and failed sets, then labels every trajectory-capability pair as NA, PRESENT, or LACKING. A capability is retained only when its contrastive gap clears δ = 0.20 and coverage clears ρ = 0.10. This filtering keeps only skills whose absence concentrates in failures.

Second, a generation agent builds one synthetic environment per retained capability. Each environment isolates a single skill while preserving the target benchmark’s tool schemas and format. Task instances are procedurally generated from random seeds, so rewards need no human labels or LLM judges.

Third, each capability gets one LoRA adapter trained with Group Relative Policy Optimization (GRPO). The base model stays frozen. Each adapter adds approximately 1.6 billion trainable parameters, or 5.3% of the backbone.

Fourth, TRACE composes the adapters into a Mixture-of-Experts model with token-level routing. Only lightweight gates are trained. At inference, each token routes top-1 to a single capability expert, allowing the model to switch specialists mid-trajectory.

On τ²-Bench (customer-service tasks), TRACE improved Qwen3-30B-A3B by 15.3 percentage points over the base model and beat the strongest external baselines (GEPA and SWE-RL) by 8.6 and 8.4 points respectively. On SWE-bench Verified, the same model gained 15 points in Pass@1. On a larger backbone, Qwen3.6-27B, TRACE reached 73.2% Pass@1 on SWE-bench Verified, according to MarkTechPost. That 27-billion-parameter open-weight model surpassed GPT-5.2-Codex (72.8%) on the public leaderboard.

The stability finding may matter more than the raw numbers. Across ten independent runs, contrastive analysis recovered the same four capability deficits: structured data reasoning, multi-step task completion, precondition verification, and tool calling precision. The failure modes are not random. They are structural, and they are consistent enough to train against.

Verifiers v1: The Plumbing Layer

Prime Intellect’s contribution targets neither the environment nor the training signal. It targets the infrastructure that connects them. NCT covered the Verifiers v1 launch on July 13. The central design decision is decomposition: split monolithic evaluation environments into three interchangeable components. A taskset defines the data, tools, and scoring. A harness solves the task and produces a rollout (it can be a ReAct loop, a CLI agent, or a custom implementation). A runtime executes the rollout locally or in a sandbox.

This matters because benchmark lock-in has been one of the quiet costs of agent RL. Teams building on SWE-bench can’t reuse their harness for customer-service evaluation without rewriting infrastructure. Teams evaluating with Codex can’t swap to a different agent framework without rebuilding the plumbing. Verifiers v1 lets any taskset run under any compatible harness, according to Prime Intellect’s technical documentation.

The implementation includes a verifiers-managed interception server that sits between the agent’s runtime and the inference server. It proxies requests, records traces, sets sampling parameters, and can rewrite tool responses to mitigate reward hacking during training. Three wire-format dialects ship out of the box: OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages.

One architectural detail stands out. Verifiers v0 used quadratic-growth prompt-completion pairs for trace storage. v1 replaces this with a linear message-graph, making long-horizon agent training tractable. For multi-step agents that run hundreds of tool calls per trajectory, the difference between quadratic and linear trace growth determines whether training is feasible at all.

Three Layers, One Stack

The timing is coincidental. The convergence is not. Each framework addresses a layer that the other two leave open.

MORPHEUS provides the environments: persistent, non-stationary worlds where agents must adapt continuously and where no fixed policy can remain optimal. Its six-metric protocol replaces single-scalar evaluation with measures of adaptation speed, forgetting, and recovery.

TRACE provides the training methodology: instead of training broadly and hoping the agent picks up what it needs, identify the specific capabilities it lacks and train each one with dedicated data and a dedicated adapter. The MoE composition then recombines them without interference.

Verifiers v1 provides the infrastructure: modular components that let researchers swap benchmarks, agent harnesses, and runtimes without rebuilding from scratch. The linear trace format and dialect-agnostic interception make it possible to plug TRACE-style training into MORPHEUS-style environments without either team knowing the other exists.

Six months ago, none of these tools existed. Teams evaluating production agents had to build their own environments, diagnose failures manually, and write evaluation infrastructure from the ground up. The research question was whether agents could pass benchmarks at all.

That question is settling. TRACE put a 27B open-weight model above GPT-5.2-Codex on SWE-bench. MORPHEUS showed that every continual RL algorithm tested sits near random performance in non-stationary enterprise simulations. The new questions are harder: Can agents adapt when the environment changes underneath them? Can training target the specific skills an agent lacks rather than flooding it with generic data? Can evaluation infrastructure keep pace with the rate at which new agent architectures ship?

Three teams, working independently, each answered one of those questions with open-source tooling on the same day. The agent RL evaluation stack now has its first generation of purpose-built infrastructure. Whether the next generation of production agents is actually good enough to benefit from it is the test that comes next.