The Celeste advantage: The harness that empowers your agents to be the best they can be

Enterprise agent adoption has surged over the past six months and is expected to keep accelerating. Governance hasn’t kept pace: Only one in five say they have the right infrastructure to support broader agentic adoption.¹ That gap carries big risks. Gartner predicts that more than 40% of agentic AI projects will be canceled in 2027 because of costs and risk control, rather than AI capabilities.

Both problems start in the infrastructure around the model, and a harness is where firms can address them.²

What is an agent harness?

An agent harness is the runtime layer that wraps around a large language model (LLM) and governs how an agent uses the LLM. It manages the context, memory, tools, and rules an agent works with.

Two ways to build an agent ecosystem

As firms move from piloting agents to running them at-scale, the foundation they choose shapes both cost and control. Here’s how two approaches to that foundation compare: 

• The standard approach: Connect an LLM to a generic agent framework
  Without an understanding of your firm’s workflows, agents will start every interaction from scratch. The result: higher costs, lower efficiency, and more iterations to produce a viable output.
• The Intapp approach: Build on the Celeste harness
  With a harness that Celeste builds automatically from your firm’s usage, the agent starts with an understanding of how your firm works. The result: lower costs, higher efficiency, and fewer iterations to create a usable output.

Where the Celeste harness fits

The Celeste harness sits between the chat application and the LLM:  

One rule keeps the harness clean: If removing any Intapp product or database would break the code, it doesn’t belong in the harness. The harness encodes your firm’s rules and context without creating dependencies on specific products or underlying data.

What the harness enables

Running a question through the harness changes the response in two ways:

1. Responses shaped by firm intelligence
   The harness tunes tool descriptions, reminders, hooks, and system guidelines to your workflows. The agent starts with a nuanced understanding of how work gets done at your firm, including which connectors and Model Context Protocols (MCPs) to surface for a deal screen, matter intake, or quality of earnings (QoE) analysis. 
2. Lower cost per answer
   The harness carries the weight of context and memory so the model doesn’t have to. Summarization and compaction that respect token budgets keep conversations from ballooning, and tool results are trimmed and reranked before they re-enter the model.

How the harness manages runtime differently

The harness is built for end-to-end workflows, with five components behind it:

• ReAct: The base loop: reason, act, observe. It owns every node and edge in a turn, including all hooks, tool call wrapping, context management, and the single point where the model is actually called. Nearly every agent in Celeste ultimately runs on top of it. 
• DeepAgent: A thin wrapper over ReAct that pre-binds planning and workspace tooling. It registers a running to-do list for multistep work. When the agent has a workspace, it adds file tools to list, read, write, edit, search, plus a code-execution tool if a sandbox is attached. It’s the default for any agent that plans ahead or needs durable files, which in practice is nearly every product agent. 
• Subagent delegation: A task tool can hand work to a specialized subagent that shares the parent agent’s workspace, state, and checkpoint, so context carries over. 
• Long-lived memory: A pluggable memory store automatically feeds episodic and procedural memory into an agent’s context. This is the same capability covered in this previous blog. 
• Open extensibility: Firms can add their own external tool servers on top of a first-party library of Intapp-built tools.

Built-in compliance and firm judgment

Every tool call, retrieval, and summary an agent makes will run through Celeste compliance checks automatically. If something fails a check, it doesn’t happen. For higher-stakes actions, the harness can pause a tool call and require human sign-off before execution. That’s what makes Celeste viable for the most regulated industries in the world.

The harness builds in firm-specific judgment at the same control point. Harness blueprints tune what the agent knows and how it behaves for your firm, layered on top of defaults grounded in decades of industry knowledge. Every firm-specific change is versioned and reviewed before it goes live, keeping specialization consistent and controlled.

Underneath, the harness guides the agent with reminders at each decision point, hooks that can adjust inputs before reaching the model, and wrappers that can stop or redirect a tool call before it runs. 

Key benefits

The Celeste harness gives your firm three advantages:

• Better responses: Vertical expertise, lower cost per answer, and fewer dropped threads on long tasks all come from running through a harness. 
• Structural compliance: Celeste enforces compliance at every step an agent takes, not as a gate at the API’s edge. 
• Compounding investment: Because the harness carries no product knowledge, new agentic features build on the same foundation and runtime instead of starting from scratch. 

Together, these advantages address the cost and risk concerns that threaten agent projects, and they give your firm a foundation to build on as agentic adoption grows. 

Sources: 

_{¹ Deloitte (2026). The state of AI in enterprise. https://www.deloitte.com/us/en/what-we-do/capabilities/applied-artificial-intelligence/content/state-of-ai-in-the-enterprise.html} 

_{² Gartner, Inc. (2025, June 25). Gartner predicts over 40% of agentic AI projects will be canceled by end of 2027 [Press release]. https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027}