Every enterprise leader we talk to is closely tracking the fundamental architectural shift in AI: from prompt engineering to context engineering, and now to harness engineering. Over the last year, most enterprises focused on optimizing prompts, selecting models, and improving the quality of context provided to those models. Those investments still matter. Better prompts and better context continue to improve outcomes.
But as frontier models commoditize and become more interchangeable, the strategic control points are shifting downstream. Model performance is no longer the primary differentiator. Critical levers such as cost control, data sovereignty, and security do not reside within the model—they reside in the execution layer.
The agent harness is that execution layer. The harness determines how context is retrieved, how tools are invoked, how pre-inference guardrails are enforced, and how tokens are consumed. For enterprises building durable AI, harness ownership is no longer an architectural afterthought. It's the cornerstone of a sovereign AI strategy.
The compounding cost of harness renting
Relying on a third-party, black-box agent harness creates architectural vulnerabilities that compound over time. Without harness ownership, your enterprise will pay for it across three critical dimensions:
- Economics: Without control over the agent loop, every task burns more tokens, more turns, and more compute budget than necessary. Competitors running cost-optimized harnesses will achieve identical outcomes at 40% to 50% lower inference spend, driving superior unit economics.
- Governance: Security, compliance, and policy guardrails remain external filters rather than native capabilities. This leaves AI systems exposed to prompt injection, unauthorized tool execution, data leakage, and auditability gaps.
- Durability: Long-running work becomes fragile, dependent on external orchestration, and unable to pause or recover cleanly from API call failures.
This gap between harness owners and harness renters will become increasingly expensive to ignore.
Gartner predicts that AI inference costs per agentic workflow will increase more than fivefold through 2028. Benchmark results demonstrate why this matters: when running the exact same underlying model, harness engineering alone can reduce token consumption by up to 75% and cut costs by more than half while achieving comparable or superior outcomes.
Same model, vastly different economics. The ability to own, tune, and control the harness creates the difference.
Introducing Tera Harness: The open enterprise harness for data work
At Teradata, we believe every enterprise must own its harness.
The agent harness governs AI economics, model choice, data sovereignty, and vendor lock-in. It's the most critical IP in the age of AI and is too strategic to become another vendor-controlled black box or an endless execution tax. That's why we built a harness for complete ownership, flexibility, and sovereignty.
Introducing Tera Harness, the execution engine behind Tera, Teradata's agentic coworker for enterprise data work. It's purpose-built for enterprise data work and grounded in five core architectural principles.
- Open by design: Tera Harness is built on Loom, Teradata's open-source enterprise agent framework, and extends the open Agent Skills standard. Customers can inspect, extend, and adapt the execution layer without vendor lock-in.
- Neutral and connected: Tera Harness is designed to integrate seamlessly into your existing enterprise stack rather than forcing workloads into a closed ecosystem. It connects natively across any model provider, Model Context Protocol (MCP) tools, line-of-business applications, relational databases, and enterprise vector stores.
- Fully customizable: Tera Harness allows administrators to configure model assignments, guardrail policies, approval workflows, context budgets, skills, and execution controls directly through the product UI.
- Purpose-built for enterprise data work: Tera Harness is engineered specifically for complex data workloads. It natively orchestrates data, models, context windows, and skills for advanced analytics and enterprise data work.
- Deploy anywhere. Stay in control: Tera Harness operates flexibly across cloud, hybrid, sovereign, and on-premises environments while supporting both open-weight and frontier models.
Tera Harness achieves higher accuracy than Claude Code at one-third the cost
Tera Harness isn't just a theoretical advantage. It has been validated against some of the industry's most demanding software, data engineering, and analytics benchmarks. Across every benchmark, Tera Harness matched or exceeded the performance of specialized tools, such as Claude Code, while delivering materially better economics.
- SWE-bench Pro: One of the industry's most demanding coding benchmarks, Tera achieved 72.3% accuracy versus 71.5% for Claude Code using the same underlying model, while delivering results at 58% lower cost ($718 vs. $1,729) and consuming 3.8× fewer tokens. Tera completed every task, while Claude Code timed out on 11 tasks.
- data-eng-bench: A benchmark created by Snowflake Labs and Bespoke Labs, Tera Harness led the benchmark on reliability with a 65.0% Pass^3 rate versus 60.0% for Claude Code, while reducing the cost per reliably delivered pipeline from $3.40 to $1.23.
- ADE-bench: A benchmark developed by dbt Labs for analytics engineering, Tera achieved 75% accuracy versus 69% for Claude Code, while lowering cost per task by 27% (~$0.19 vs. $0.26). It also executed tasks 35% faster while consuming 3.3× fewer tokens.
These aren't just incremental gains. They're proof that an execution-centric, cost-optimized harness can transform the same model into a more accurate, reliable, and significantly more cost-effective system. Read our companion benchmark report for more details.
Harness design for long-running enterprise data work
The mainstream narrative on harness design is converging on two dominant approaches: closed platform harnesses and open meta-harnesses. Both approaches focus primarily on orchestration, model routing, and agent coordination.
Yet in our conversations with enterprise customers, particularly in highly regulated industries, a different reality keeps surfacing. Orchestration alone fails at scale. Real enterprise value demands a neutral, execution-centric harness grounded in three critical capabilities: reliability, cost-predictability, and durability for long-running business processes.
We designed Tera Harness to solve this fundamental gap. Most harnesses are designed for orchestration. Tera Harness is built for execution. Orchestration determines what should happen next. Execution ensures the work actually gets done—reliably, cost-effectively, and at enterprise scale.
1. Cost-optimized agent loop
Standard agents depend on repeated LLM reasoning at every step of the loop, re-reading context, making serial tool calls, and creating expensive trial-and-error execution. Tera addresses this through:
- Cost-aware execution: Chooses the cheapest reliable engine for each sub-task—frontier LLMs, SLMs, or native in-database compute.
- Cache-aware harness: Keeps the full session history cached so the model never reprocesses what it already knows.
- Batched tool calls: Completes more work per model interaction by executing reads, writes, and queries in batches.
- Pattern-guided execution: Applies 84 proven execution patterns pre-inference, so agents start with an execution plan.
2. Loop-embedded guardrails
Standard agents rely on self-policing or external filters, leaving enterprise systems exposed to prompt injection, destructive database operations, and unauthorized schema changes. Tera addresses this by embedding governance directly into the execution loop.
- Pre-inference guardrails: Intercepts high-risk actions before execution, blocking unauthorized writes, destructive SQL, and schema changes.
- Customizable governance policies: Allows administrators to define which operations are allowed, denied, or routed for approval.
- Admin-configurable harness: Gives teams control over model assignments, guardrail thresholds, skill activation, and context budgets through product UI.
3. Go-native performance core
Standard agent frameworks rely on heavy Python runtimes that run each agent as a separate process. At enterprise scale, infrastructure costs and latency spiral with every agent added. Tera addresses this through:
- Go language implementation: Lightweight Go concurrency enables thousands of parallel operations with minimal memory and CPU overhead.
- gRPC service layer: Replaces REST APIs with a high-performance, contract-first communication layer that minimizes latency and execution overhead.
- Protocol buffer contracts: Binary encoding replaces text-based JSON, eliminating repeated string marshalling and reducing communication overhead on every tool call.
4. Durable agentic execution
Most agent frameworks depend on external orchestration layers to survive failures and maintain state, while continuing to consume cloud resources during long-running waits. Tera addresses this through:
- Built-in state management and checkpointing: Embeds state management and checkpointing directly into the runtime, eliminating the need for a separate orchestration layer for durability and recovery.
- Zero-cost wait: Allows agents to pause while awaiting approvals or external events without continuously consuming compute resources.
- Automatic recovery: Restores agent state and execution context after failures, restarts, or infrastructure interruptions without losing work.
Measuring harness ROI: Cost per reliable outcome
When evaluating agentic systems, enterprise leaders often focus on model metrics such as cost per million tokens, latency per turn, or benchmark scores. These metrics miss the point. What ultimately matters is simple: what does it cost to complete a business-critical workflow correctly, reliably, and repeatedly?
Consider two systems running the exact same model.
- An orchestration-centric harness operates open-loop and is designed primarily to coordinate agents, models, and tools. It relies on the model to repeatedly discover context, determine execution paths, and recover from mistakes. As workflows become more complex, context gets reprocessed, work gets repeated, and errors propagate downstream. The result is higher token consumption, increased compute costs, and a greater risk of hallucinated, incomplete, or incorrect outcomes.
- An execution-centric harness operates closed-loop and is purpose-built for enterprise data work. It continuously validates execution state while coordinating specialized skills, data tools, and workflow patterns throughout the task. Instead of relying on the model to repeatedly discover context and determine the next step, the harness provides guided execution that reduces token consumption, prevents downstream errors, and improves outcome reliability.
This is why enterprise AI must move beyond cost per token and even cost per outcome. The true measure of AI ROI should be cost per reliable outcome.
What enterprise teams can deliver with Tera Harness
In customer conversations and early proofs of concept, we're seeing demand for a new class of AI workloads that existing agent frameworks struggle to handle. Banks are exploring agents that can conduct multiday AML investigations across millions of transactions. Healthcare players are evaluating agents to detect fraud and anomalies across massive claims and patient datasets. Life sciences teams are piloting multi-horizon trial synthesis across fragmented data warehouses.
Single-threaded Python wrappers and open-loop orchestrators collapse under these workloads. As workflows grow more complex, they lose state during transient failures, burn tokens recovering from unverified errors, and create gaps in governance and auditability.
Purpose-built in Go for high-concurrency execution, Tera Harness combines durable state persistence with native enterprise data skills to deliver the reliability, governance, and cost predictability that regulated enterprises require. This translates into measurable business outcomes:
- 60% fewer model turns
- 74% lower token consumption
- 58% reduction in execution costs
- 42% faster task completion
Let's talk execution
Discover how Tera Harness delivers reliable outcomes with lower cost, stronger governance, and greater operational control. Reach out to us to get early access and explore how Tera Harness delivers true deployment choice across Teradata Cloud and Factory.