Get Started
Topic · 100 recaps
AI Agents
Research on AI systems that plan, use tools, and act autonomously across multi-step tasks — from chat-only assistants to embodied robotic agents.
Play all
...
Posts
Questions
Home
Topics
Search
Library
Questions researchers are working on
Follow a question through Rcap’s explanations and the latest papers addressing it.
Search
Can chain-of-thought monitoring detect preferences received through tools or inferred from raw artifacts?
Chain-of-thought monitoring assumes that a model’s reasoning trace reveals the information influencing its answer. Preferences delivered through tool returns or inferred from unprocessed artifacts may affect answers without being clearly verbalized in the trace.
Can language agents maintain hidden state consistently across dialogue branches using only public conversation history?
A chat interface exposes conversation history but provides no separate channel for state that must remain hidden. When dialogue branches, the agent must preserve the same secret and answer consistently without revealing or reconstructing it from public text.
How can a single search agent improve multi-hop web research without sub-agents or test-time verification?
Multi-hop web research requires an agent to connect evidence across several pages while retaining the information needed for later steps. Long search trajectories can overwhelm the available context, making both training and reliable performance difficult without additional agents or verification passes.
How can adaptive trading agents be stress-tested across alternative futures when returns hide state and execution failures?
Historical backtests expose an adaptive trading agent to a single realized market path and cannot rule out historical contamination. Terminal portfolio returns can also look successful when internal state has degraded or reported decisions no longer match executed holdings.
How can agent decisions be reconstructed for auditing and controlled replay when tool state and authorization context are missing?
An agent’s final output does not reveal the evidence, tool state, authorization, or action path that produced a committed decision. Missing or unobserved execution state makes it difficult to distinguish a faithfully replayed decision from an unexplained divergence.
How can agent harnesses adapt across tasks and models without manual redesign?
Agent harnesses coordinate memory, planning, actions, and tools, but they are often manually designed for individual tasks. This makes it difficult to reuse them across models or adapt them reliably as task requirements change.
How can agent runtimes avoid context poisoning and latency from growing histories during long-horizon skill execution?
Long-running agents often append observations, actions, and intermediate reasoning to their conversation history. As that history grows, execution becomes slower and earlier context can interfere with later decisions.
How can agent skills support reliable procedural execution while making failures easy to diagnose and repair?
When skills are written as free-form prose, agents must repeatedly infer procedural steps, code, commands, and tool calls, which can reduce reliability on implementation-heavy tasks. The same representation makes it difficult to locate failures and safely improve domain-specific procedures.
How can agentic benchmarks be compared and reused across complex environments and bespoke agent integrations?
Agentic benchmarks depend on complex environments and bespoke integrations, making them difficult to run consistently across agents. This limits reliable comparison and broad reuse of benchmark results.
How can agentic vision-language models acquire and use necessary external evidence without redundant tool calls?
Complex image-grounded questions may require visual details or external knowledge unavailable in the initial input. Models may pursue irrelevant evidence or fail to extract useful information from tool outputs, while final-answer supervision does not clearly teach effective evidence acquisition.
How can agents adapt as their tool, skill, and specialist-agent harness evolves without losing existing capabilities?
Adding tools, reusable skills, or specialist agents can change what an agent observes and can do, disrupting tasks it previously solved. The agent must accommodate new capabilities while preserving useful prior behavior.
How can agents choose task-dependent world-model rollout horizons for effective multi-step planning?
Learned world models let agents imagine future environmental states without interacting with real environments. Single-step or fixed-horizon imagination may provide insufficient foresight when tasks require different amounts of planning and progress changes over time.
How can agents find people across cameras from vague witness clues under spatial-temporal and turn constraints?
Witness accounts may be partial or ambiguous, while relevant observations are distributed across camera locations and time. An agent must choose questions and searches before its interaction budget runs out.
How can agents make their intentions and internal perspective interpretable to observers?
Observers must infer an agent’s hidden intentions and perspective from its behavior and any explanations it provides. This is difficult when the agent’s internal reasoning is not directly accessible.
How can AI agent harnesses prevent trusted plugin updates from triggering host-privileged attacker commands through lifecycle hooks?
Lifecycle hooks can bind shell commands to routine agent events and execute them with host privileges, including at times the model may not observe. A malicious update can therefore transform benign plugin configuration into host-side behavior without obvious agent involvement.
How can AI agents adapt execution routes as runtime evidence invalidates their planned continuation?
Agents often commit to a multi-step execution route before intermediate outcomes are known. When those outcomes change what should happen next, stale actions or broad replanning can waste computation and discard useful progress.
How can AI agents adapt to each user’s evolving quality criteria on open-ended tasks?
On open-ended writing and visual-creation tasks, users often judge outputs by standards they cannot fully articulate in advance. These standards may emerge and shift during interaction, leaving broadly capable agents unable to consistently meet an individual’s professional expectations.
How can AI agents make controllable slide edits while preserving layout, styling, structure, and native editability?
Editing a slide requires more than producing a plausible rendered image. Weak representations can fragment coherent visual units, disrupt layout and styling, or rasterize content that should remain editable.
How can AI agents onboard and control unfamiliar hardware across platforms without bespoke integrations?
Integrating each new device commonly requires manual driver selection, dependency setup, interface design, and deployment work. Platform differences make those steps difficult to reproduce and maintain across hosts.
How can AI generate attractive graphics with accurate text and editable layers?
Bitmap generation often flattens designs, making text unreliable and later edits difficult. Code-based generation preserves structure but can struggle with aesthetic judgment and complex visual assets.
Previous
1 / 10
Next