When an AI agent spends money, the bill is not determined only by the model answering a question. It is also shaped by the software deciding what the model should see, which tools it should call, how often it should repeat information and where its unfinished work should live. TrueFoundry’s open source TrueForge aims to make that layer cheaper and more controllable, while raising a larger question for enterprise buyers: is the future of AI agents managed convenience, or infrastructure that companies can own?
For many businesses, the first encounter with an AI agent is deceptively simple. An employee asks the system to investigate a customer problem, update a record, search internal documents and create a follow-up ticket. The agent appears to do the work in one continuous exchange.
Behind that exchange, however, the system may be performing dozens of smaller operations. It may retrieve tool instructions, inspect application data, send results back to a language model, decide on the next action, open a secure environment for code execution and repeat the process several times. Each step can consume tokens, computing resources and time. A task that looks like one request to a user can become a long chain of model calls and tool interactions.
TrueFoundry, an AI infrastructure company, is betting that enterprises will increasingly want control over that chain. On August 19, VentureBeat reported that the company had released TrueForge, an MIT-licensed agent harness designed to coordinate models, tools, context, sandboxes and persistent work. The company describes it as vendor-neutral and suitable for local deployments as well as shared production environments.
The release is significant not simply because another AI developer tool has entered the market. It reflects a shift in where companies are looking for savings and leverage. In the early years of generative AI, buyers focused heavily on the price and quality of the underlying model. As agents take on longer, more complicated assignments, the runtime around the model is becoming just as important.
The expensive part may be everything around the model
A language model does not work in isolation when it is used as an agent. It needs information about the available tools. It needs the results of previous actions. It may need access to a browser, a database, a file system or a code execution environment. It must also retain enough history to understand what has already happened.
The more information that is sent into a model, the more expensive and potentially slower each call can become. This is particularly true for agents that operate over many steps. A tool may return a large collection of records even though the agent needs only a small portion of them. A runtime may repeatedly include the same instructions in a growing conversation. A system may provide descriptions of every available tool, even when the agent needs only one.
These inefficiencies can remain invisible during a small pilot. A team may run a few successful demonstrations and conclude that an agent is affordable. The economics change when thousands of employees begin assigning work to the system, or when an agent is asked to complete a process that requires sustained interaction with several enterprise applications.
TrueForge is designed around the idea that the runtime should actively manage this burden. According to TrueFoundry, it can load Model Context Protocol tool schemas only when they are needed, rather than placing every possible tool description into the model’s working context. It can delegate isolated pieces of work to subagents, store large results in files rather than keeping them in the active conversation, and process structured data in code instead of asking the model to reason through every item.
The harness can also compact long conversations and provision a sandbox only when the agent needs to manipulate files or execute code. In practical terms, the system is trying to behave less like a person carrying every document in a meeting room and more like an organized assistant that retrieves the relevant folder when required.
That distinction matters because context is not free. It can influence cost, latency and accuracy at the same time. A model that receives too much information may spend more money and still make a worse decision because the important detail is buried inside irrelevant material.
TrueFoundry’s benchmark claim
TrueFoundry’s most prominent claim concerns a benchmark involving multi-step work across customer relationship management, issue tracking and document management systems. In a DevRev Enterprise-Bench test, the company said TrueForge paired with GLM-5.2 completed 11 of 14 tasks for $2.90. Anthropic’s Claude Managed Agents, using Claude Opus 4.8, completed the same number of tasks for $11.80, according to the company’s figures.
The headline comparison suggests a cost reduction of roughly 75%. TrueFoundry also reported that when Opus 4.8 was used in both systems, TrueForge completed the benchmark for $8.50, compared with $11.80 for Claude Managed Agents. That represents a saving of about 30%.
Those numbers are noteworthy, but they should be read as vendor-reported benchmark results rather than a universal price promise. Agent costs depend on the model, the task design, the number of tool calls, the size of returned data, the reliability of external applications and the policies used by the runtime. A different collection of tasks could produce a different ranking.
The completion rate also deserves attention. The reported figure was 11 out of 14 tasks, not a perfect score. That does not invalidate the comparison, but it puts the cost claims in context. A cheaper system is not necessarily better if it requires a person to repair failed workflows, repeat tasks or inspect every result. The relevant measure for a company is not simply the cost of a successful model call. It is the total cost of getting dependable work completed.
That includes human review, exception handling, security checks, maintenance and the opportunity cost of engineers who must tune the system. It also includes the consequences of a wrong action. An inaccurate summary is inconvenient. An incorrect update to a customer account, an erroneous ticket closure or an unauthorized file change can be considerably more expensive.
The benchmark therefore points to a meaningful question rather than settling one. Can a carefully designed runtime lower the cost of useful agent work without introducing unacceptable reliability problems?
Context engineering becomes a competitive discipline
The techniques described by TrueFoundry belong to a broader movement toward what might be called context engineering. The phrase refers to the deliberate management of the information an AI system receives, rather than treating the model’s context window as an unlimited storage area.
The distinction is becoming important as companies move from chatbots to agents. A chatbot can often answer a question from a relatively stable prompt. An agent must navigate uncertainty. It decides what to do next, interprets the response from an external system and chooses whether another action is necessary.
Every decision creates an opportunity for waste. An agent can call the wrong tool. It can ask a model to inspect data that software could filter more accurately. It can preserve an entire conversation when a short summary would be sufficient. It can expose a long list of capabilities that makes it harder to select the right one.
TrueForge’s approach is to push some of that work into the runtime. Structured results can be processed programmatically. Large outputs can be placed in files and referenced when needed. Subagents can handle focused assignments without forcing the main agent to carry every detail. Context can be compressed as a task continues.
This resembles a basic principle of operations management: do not move more material through a production line than the next station requires. In an AI system, unnecessary context is the material. The runtime becomes the logistics department that decides what should move, what should be stored and what can be discarded.
There are trade-offs. Summarizing a conversation can remove a detail that later turns out to matter. Moving information into files can make the system more efficient but may complicate access control and auditing. Delegating work to subagents can reduce the main agent’s burden while making the overall process harder to observe. A sandbox can improve safety, but it also introduces configuration and monitoring requirements.
The challenge is that efficiency techniques can change behavior. An enterprise cannot assume that a lower token bill represents a pure improvement. It must test whether the agent remains understandable, recoverable and consistent when the runtime manages information more aggressively.
Open source at the execution layer
TrueForge’s MIT license gives organizations broad freedom to inspect, modify and deploy the harness. That freedom may appeal to companies that do not want their agent architecture tied too closely to one model provider or one cloud platform.
Vendor neutrality is especially valuable in a market where model capabilities and pricing are changing quickly. A company might use one model for complex reasoning, another for high-volume classification and a third for tasks that must run locally. An independent harness can offer a common control layer while allowing those choices to evolve.
Local deployment can also matter for organizations with strict data rules, sensitive intellectual property or requirements to keep certain operations within a specific jurisdiction. Shared production deployments, meanwhile, can provide centralized administration for teams that do not want every department building its own agent infrastructure.
Yet open source does not automatically mean low total cost. Running a production agent platform requires engineers, security specialists and operations staff. Someone must manage upgrades, monitor failures, control credentials and determine what the system is allowed to do. A company that saves on model and runtime fees may spend more on internal expertise.
This is where TrueFoundry’s commercial strategy becomes important. The company positions its AI Gateway as the place for enterprise controls, including centralized access to models and MCP tools, credentials, identity, permissions, budgets, observability and failover.
The pattern is familiar across enterprise software. A company releases a foundational layer that developers can adopt freely, then sells the management systems that make the technology easier to govern at scale. Open source accelerates experimentation and builds a community. Commercial controls address the concerns of security, finance and compliance departments.
That arrangement could work well for customers, but it also means the distinction between independence and dependence may become less clear over time. An organization might begin with a self-hosted TrueForge deployment and later adopt the company’s gateway for convenience. The open execution layer would remain portable in theory, while operational habits and governance policies could gradually tie the customer to one provider.
The governance problem behind cheaper agents
The enterprise argument for an agent harness is not limited to cost. As agents gain access to business systems, the runtime becomes part of the organization’s control structure.
A model may be capable of calling a tool, but the runtime determines whether it should be allowed to do so. It can enforce permissions, limit budgets, record activity and route requests through approved services. It can help answer questions such as which employee authorized an action, which model made the decision and what data was available at the time.
These questions will become more important as regulators, auditors and corporate boards examine automated decision-making. A business may need to demonstrate that an agent did not access information outside its role or that a sensitive action required human approval. It may need to reconstruct an incident after a tool returned unexpected data.
Managed agent products offer convenience because many of these functions can be bundled into a provider’s platform. The customer gives up some control in exchange for a shorter path to deployment. An open harness gives the customer more freedom, but also more responsibility for building and maintaining the surrounding safeguards.
There is no universal answer. A heavily regulated bank may prefer centralized controls and a narrow set of approved models. A software company with a strong platform engineering team may value the ability to tune every part of the workflow. A smaller business may choose a managed service because it cannot justify maintaining its own runtime.
The important change is that the choice is no longer only between models. Companies are choosing how much of the agent stack they want to own.
Managed convenience versus infrastructure ownership
Anthropic and other model providers have an obvious advantage in this contest. Managed agents can be easier to start with because the provider controls the model, the runtime and much of the deployment experience. Customers may receive integrated tool access, scaling and support without assembling a platform from separate components.
That convenience has a real economic value. Engineering time is expensive, and delays can erase the savings from a lower per-task bill. A provider that offers a dependable managed system may be able to charge more while still delivering a lower overall cost of ownership.
The risk for customers is concentration. When the model, runtime, tools and operational controls come from one supplier, switching can become difficult. A company may have to rewrite workflows, retrain staff and recreate monitoring systems if prices rise or performance changes. It may also have less visibility into how the provider manages context and decides which operations are performed.
TrueForge represents the alternative model. The customer can choose the model, inspect the harness and decide how the system is deployed. That may produce better economics for companies with the right technical capabilities. It may also enable more specialized agent architectures, particularly in environments where data placement and tool permissions are critical.
But portability is valuable only when it is practical. An agent built around provider-specific features may still be difficult to move, even if the underlying harness is open source. Tool definitions, authentication systems, evaluation methods and data formats can all create friction. True vendor neutrality requires more than accepting multiple model endpoints. It requires a durable ecosystem around the runtime.
What companies should test before believing the savings
Organizations considering TrueForge or a similar system should begin with their own work patterns rather than a headline benchmark. A useful evaluation would measure the full cost of completing representative tasks, including model calls, infrastructure, engineering time and human intervention.
The test should include straightforward successes and difficult cases. It should examine how often the agent calls tools unnecessarily, how it behaves when a system returns incomplete data and how much context it carries after a long sequence of actions. Teams should compare not only the final price but also latency, error rates, recovery time and the effort required to audit results.
They should also test different models. TrueFoundry’s comparison suggests that the harness itself can affect cost even when the same model is used. That is a significant possibility, because it would mean runtime design can produce savings without requiring a company to accept a weaker model. Still, the result needs to be reproduced under realistic workloads.
Security testing is equally important. A sandbox should be evaluated against attempts to access unauthorized files or systems. Tool permissions should be checked at the user and service level. Logs should make it possible to understand why an agent took an action, not merely record that it happened.
Finally, companies should ask whether their teams want to operate an agent platform. Open source can provide strategic control, but control brings duties. The organization must be prepared to patch vulnerabilities, evaluate new models, manage changing protocols and investigate failures.
The runtime may become the real battleground
The AI industry has spent much of its energy competing over models. That contest will continue, but agents are shifting part of the struggle to the layer that directs model behavior.
The runtime controls how much a model sees, how often it is called, which tools it can use and how its work is preserved. Those decisions affect cost and reliability in ways that may be invisible to an end user. They also shape whether a company can understand and govern its automation.
TrueForge’s release arrives at a moment when businesses are moving from demonstrations to operational systems. The question is no longer whether an agent can complete one impressive task. It is whether thousands of tasks can be completed at a predictable price, under defined permissions, with a clear record of what happened.
TrueFoundry may be right that a well designed harness can reduce the cost of agentic work by a substantial margin. The company’s benchmark claims are a useful signal that the runtime deserves as much scrutiny as the model. But the market will ultimately judge TrueForge on more than its token efficiency.
It will be judged by whether agents remain dependable when context is compressed, whether failures can be diagnosed, whether organizations can move between models and whether open infrastructure creates genuine control rather than another route into platform dependence.
For the people who use these systems, the difference will appear in ordinary moments: a customer issue resolved without a mistake, a report assembled without hours of manual checking or a sensitive file left untouched because the agent understood its limits. The future of enterprise AI may depend less on how impressive an agent sounds than on how carefully the invisible machinery behind it manages the work.