By enabling the harness, we move one step closer to achieving dynamic context discovery.
Context engineering is the process of determining what information enters an agent's limited context window. Context is a finite resource with diminishing returns. Humans have a limited working memory, similar to that of Large Language Models; however, instead of working memory, LLMs operate within an attention budget. Every additional word, or in LLM terms, a token, consumes a portion of that budget. As context grows, the model must divide its attention across more information, making curation of what enters that window essential for reasoning.
To address this challenge, Relay leverages three context systems: Engagement History, Firm Methodology, and Chat History. Together, these systems ensure that relevant information is surfaced to the agent when needed, maximising signal while minimising noise.
Through dynamic context discovery, Relay builds an understanding of your clients and your firm's methodology. This enables Relay to uphold your company's standards and processes while ensuring the expectations and requirements of each client are met.
Engagement History
Engagement graph loaded.
High Level Audit Knowledge Base Example
Relay enables firms to upload all relevant engagement information, including prior engagements, financial documents, and supporting evidence. Through our architecture, this information is organised into a consolidated knowledge network.
This network connects related information across multiple sources, enabling agents to trace relationships, perform multi-step reasoning, and retrieve the information most relevant to a given task. As a result, agents can optimise the use of their context window without sacrificing output quality.
The network evolves alongside the engagement. As new information is introduced, it is embedded into the knowledge graph, ensuring that the underlying context remains current and relevant.
By maintaining an up-to-date understanding of the engagement, agents produce more accurate outputs while reducing the number of iterations required to reach a final result.
Firm Methodology
What are skills?
A skill is a structured, reusable package of procedural knowledge that tells an agent, step by step, how to perform a specific task. Agent skills allow any compatible agent to dynamically access targeted expertise on demand.
Before skills, users relied on inefficient workarounds: long, detailed prompts or bloated system messages that attempted to teach the agent everything upfront. These approaches consume context-window tokens and reduce flexibility.
Skills change that. Firms can build custom skills tailored to their audit methodology - checklists, data-extraction routines, testing procedures, and decision rules - that the agent can invoke when needed. Skills are reusable, composable, and can be versioned or fine-tuned.
Agents may also be granted controlled autonomy to explore and reason through audit tasks. When an agent discovers a useful routine during an engagement, it can encapsulate that routine as a new skill for later review and refinement.
Benefits for Audit Teams
- Efficiency: Repeatable procedures reduce time spent drafting instructions.
- Consistency: The same methodology is applied across engagements and auditors.
- Auditability: Skills produce standardised outputs and documented rationale, simplifying review.
- Continuous Improvement: Teams can version and refine skills based on lessons learned or evolving standards.
- Controlled Autonomy: Agents can generate candidate skills during an engagement, but firms retain governance by reviewing, approving, and integrating those skills into the official library.
Chat History
An agent running in a loop continuously generates new information, much of which may become relevant in later turns. Instead of forcing the agent to repeat work it has already done, chat history allows it to retain and refine important context over time.
As the conversation evolves, earlier prompts often become less relevant than the most recent instructions, corrections, and decisions. This is why chat history is so important: it helps the agent prioritise what matters now while still preserving the critical details from earlier in the engagement. In other words, the agent is not starting from scratch each time - it is building on an active thread of context.
A useful pattern here is agentic memory. In this approach, the agent periodically writes decisions to a knowledge base outside the main context window and knowledge graph, then pulls them back in when needed. This creates persistent memory with low overhead and helps the agent maintain continuity across longer or more complex tasks.
You can think of it as a running project log. Just as a developer may maintain a NOTES.md file or a task list to track progress, an AI agent can store decisions, open items, assumptions, and dependencies so they are not lost as the conversation grows.
This becomes especially valuable when the agent is working through multi-step audit tasks, where earlier conclusions often affect later testing, review, or reporting.
Chat history also helps tie together other important layers of context, such as firm methodology and engagement history. For example, if the firm prefers a specific review process or documentation style, the agent can apply that preference consistently. If a prior engagement identified recurring risks or client-specific issues, those can be surfaced again without the user needing to repeat them.
This is where dynamic context becomes powerful. The agent is not only responding to the latest message - it is continuously updating its understanding based on what has already happened, what has changed, and what still needs attention. That makes the interaction more efficient, more accurate, and more aligned with how real audit work is performed.
At Relay, we believe that as models continue to improve, the challenge of maintaining coherence across extended interactions will remain central to building more effective agents.