Agents grow up: closed loops, self-driving imaging, trackable skills
InJePA pairs goal-conditioned and action-conditioned world models for intention-first visual navigation; NeuroPilot closes the loop on neural imaging with an LLM orchestrator; TrackRL brings evidence-gated lifecycle management to agentic reinforcement learning.
- In a preprint, InJePA uses a goal-conditioned JEPA to propose latent intentions and an action-conditioned JEPA to predict future latent states, enabling closed-loop navigation by feeding predicted consequences back into the next intention proposal and action inference. Self-reported m:…3145
- In a preprint, NeuroPilot uses a multimodal LLM to coordinate computational imaging modules for closed-loop neural data processing, restores motion-degraded recordings with consistent neural signal recovery in deep-brain and freely moving settings, and enables neuronal population tracking across spinal stimulation and 24-hour mPFC recordings aligned with sleep-wake behavior. Self-reported m:…2944
- In a preprint, TrackRL uses online randomized controlled experiments within GRPO groups to isolate individual skill contributions, formulates skill lifecycle management as an evidence-gated state-transition process, and outperforms recent skill-augmented RL baselines in both library-assisted and skill-free evaluation. Self-reported m:…6583
Intentions first, then actions
Three recent preprints give agentic systems the same thing in three domains: a closed loop. Navigation, neural imaging, and skill management each get infrastructure that feeds consequences back into decisions. The first is InJePA — intention-first latent planning with coupled JEPA world models for visual navigation. A goal-conditioned JEPA proposes latent intentions; an action-conditioned JEPA predicts future latent states. The loop closes by feeding predicted consequences back into the next intention proposal and action inference — closed-loop navigation, not open-loop rollout. The architecture separates the what (intention) from the how (action), which is how plans survive contact with the environment. m:…3145
The bet — reasoning: navigation fails when the world model and the policy disagree about what is happening, and coupling two world models around intentions keeps them honest with each other. Whether it scales beyond navigation tasks is the open question, but the pattern — predict consequences, feed them back, replan — is the oldest control idea there is, now wearing a world-model costume. m:…3145
Intention-first planning has a human resonance — reasoning: people navigate by deciding where to go, then figuring out how. Agents that plan the same way are easier to debug, because you can inspect the intention separately from the action that failed. That debuggability is infrastructure too, the kind executives should price into build-versus-buy decisions. m:…3145
The imaging agent that watches the brain
NeuroPilot applies the same closed-loop instinct to neural imaging. A multimodal LLM coordinates computational imaging modules for closed-loop neural data processing: it restores motion-degraded recordings and supports consistent neural signal recovery in deep-brain and freely moving settings. The agent does not just analyze a finished recording; it participates in the acquisition loop, which is what makes it an agent rather than a pipeline. m:…2944
The platform enables neuronal population tracking across spinal stimulation and 24-hour mPFC recordings aligned with sleep-wake behavior — longitudinal, intravital, across scales and modalities. That is the use case that justifies agentic imaging — reasoning: no human watches 24 hours of neural video; an agent that tracks populations through it turns a data deluge into a dataset. m:…2944
The pattern across InJePA and NeuroPilot — reasoning: the LLM is the coordinator, not the sensor or the actor. It proposes intentions and routes work to specialized modules. That division of labor is also the safety story: the modules are testable, the orchestration is auditable, and the failure modes live in named components rather than in one inscrutable forward pass. m:…2944
Skills with a paper trail
TrackRL tackles the other end of agentic maturity: skill management. It uses online randomized controlled experiments within GRPO groups to isolate individual skill contributions — causal attribution for skills, not vibes. Skill lifecycle management is formulated as an evidence-gated state-transition process: skills earn their place in the library, and the gates are evidence, not enthusiasm. And it works: TrackRL outperforms recent skill-augmented RL baselines in both library-assisted and skill-free evaluation, which means the skills transfer rather than merely decorating the training run. m:…6583
The unifying theme — reasoning: agents are getting infrastructure. Closed-loop navigation, orchestrated imaging, evidence-gated skills — each paper replaces an open loop with a closed one. The agent stops being a demo and starts being a system when someone builds the plumbing, and this week three different teams laid pipe. m:…6583
What is still unresolved
- Whether InJePA's intention-first planning transfers beyond visual navigation tasks.
- Prospective validation of NeuroPilot's closed-loop processing in live experiments.
- Whether TrackRL's skill attribution holds outside GRPO-based training.
- The compute cost of coupled world-model architectures at deployment scale.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.