Planted websites inject false context into the agent's reasoning chain via a triggered search, poisoning the context window used to rank options.
Injected fabricated context (planted websites associating 'blizyester' with dementia) is incorporated into the agent's decision context, demonstrating how external context pollution shapes agentic judgment.
Providing agents with more complete user context is shown to paradoxically increase execution of unacknowledged proposals, highlighting how richer context can undermine rather than support appropriate action scoping.
AirPlay screen mirroring provides the vision model with live visual context of the iPhone display, grounding agent actions in real UI state.
Opening lexicons and identity tags serve as context signals that the revised agent uses to select between a relational or task-execution reply mode.
Users' evolving research interests are inferred from their annotation and reading traces in the document, not from explicit chat prompts, making implicit intent legible to the agent.
Recalibration pillar addresses how users update their trust, expectations, and mental models in response to ongoing agent behavior.
The study shows that conversational history alone is insufficient to anchor agent identity; system-level context (system prompt) must be continuously re-injected to maintain persona enactment across resumed turns.
The hotline's pre-existing cultural legitimacy (Santa Claus framing) acts as ambient context that overrides persona authority effects, demonstrating how deployment context shapes agent influence.
Linked artifacts (outlines, worldbuilding, narrative graphs) serve as persistent structured context that grounds both agent execution and author guidance.
The research shows that competing referents in visual context increase errors without raising elicited uncertainty, meaning agents fail to use contextual richness to self-calibrate.
Agents monitor the evolving document as live context to ground their contributions, tying agent behavior directly to artifact state.
A language-model agent inspects live game-state information through a controlled interface to maintain situational context before deciding speech and actions.
Agents infer user wealth from ambient personal data (e.g., unrelated emails) and use it to steer recommendations even when not instructed to consider finances.
The adaptive, module-organised graph acts as a persistent shared context layer that anchors developer understanding across cumulative coding tasks.
Recurring guidance is turned into reusable assets, externalizing context that would otherwise be re-specified at each delegation.
Sponsorship disclosures are received and evaluated by the agent rather than surfaced to the consumer, making the agent's reasoning context — including how it weighs paid placements — invisible to the user it ostensibly serves.
Normative signals communicated by AI group members function as contextual framing that shifts human decision-making, illustrating how agent-set context influences downstream human choices.
The Channel Decoupling Attack splits extraction requests across independent channels, exploiting the agent's inability to correlate context across sessions — a structural vulnerability in how agents manage conversational context.
Sustained novelty is associated with users supplying broad and distinctive contextual material, making context engineering a primary lever for output quality.
Appleton argues that 'thicker interfaces' must carry richer context so agents understand the intent behind a plan, not just its surface steps.
Cross-client context injection through unprotected context identifiers allows adversaries to corrupt shared context across agent sessions.
The attack exploits the agent's context ingestion of untrusted external images to inject malicious tool calls without user authorization.
Portal selectively compresses or filters what context is fed to Claude Code, reducing token consumption by 90% while preserving task effectiveness.
The paper finds that attributes displayed on a results page matter more than placement, pointing to context engineering — what information is surfaced and how — as the critical design lever for agentic search.
The plugins/documents folder contains skill manifests that tell Codex how to locate and invoke bundled native binaries, providing the agent with structured context about available local tools.
Vercel publishes a single public URL (design.md) that any agent can load to receive brand, layout, and copywriting guidance regardless of the environment it runs in.
Karpathy describes using voice input to provide rich, unstructured context to an LLM before it performs work, reducing the friction of expressing intent.
Work state is stored as auditable, natural-language-readable objects so neither the human nor the agent loses context when leadership of the task alternates.
The proposal distinguishes elicitation for non-sensitive data (existing spec) from a new mode for sensitive context like API keys and credentials, adding a security-tiered context-gathering pattern.
The research demonstrates that attribute information displayed on a results page — not positional context — is the dominant signal AI agents use to make purchase decisions.
The classifier operates as a separate browser-based agent, maintaining independence from the conversational AI to preserve user privacy and avoid conflict-of-interest in its judgments.
MemUse introduces 'user-cued memory moments' as a benchmark for detecting when prior context is relevant and integrating it naturally, rather than treating memory as a retrieval task.
Code Mode exposes ~2,500 Cloudflare API endpoints in ~1,000 tokens by having the model generate code rather than enumerating tools, radically compressing context requirements.
The core thesis is that every token in the orchestrator's context competes for its attention, making context protection the primary justification for subagent use.