Long-Horizon Agents: From Order-Takers to Outcome Owners

Not every employee is a superstar. In large organizations, you will encounter order takers. Despite targeted coaching, repeated instruction, and close evaluation, order takers require constant supervision, learn slowly, and fail to internalize the goals (org, team, and personal) that they’re meant to carry to their daily tasks. Powered by language models and the harnesses around them, AI agents are touted as a way to replace order takers. Companies like Oracle, GitLab, and Elastic are enacting huge AI-related layoffs and the numbers just keep growing.

But despite the promises of a pure AI utopia, we still can’t trust agents to work towards goals. Most agents require an order-taker’s level of supervision, coaxing, and constant prompt-response. A human manager must still own a goal, break it into tasks, give the models context for each task, and triple check outputs and next steps. Even when an agent uses tools and skills across multiple tasks, a human for the most part, is still responsible for goals and outcomes.

From Order Takers to Outcome Owners

A task has a clear input and output. A tool can help complete it. But intelligent humans are not hired simply to complete a collection of tasks. They are given an area of responsibility and expected to deliver against goals or outcomes.

Take a software engineer. Refactoring a module or fixing compile-time errors are tasks. Making sure the backend scales with seasonal traffic patterns while minimizing cloud spend is an outcome. The engineer needs to understand the architecture, decide what needs to change (if anything), weigh tradeoffs, sequence the work, then monitor outcomes and adjust the plan when they find something unexpected. To understand seasonality, they might study historical trends or ask others in the organization. As new patterns emerge, they update their understanding and remember it to avoid future pain.

Consider a sales rep. Writing a follow-up email after a sales call and logging it in the CRM are tasks. Generating $500K in qualified pipeline is an outcome. The rep needs to understand the customer needs and product value, identify and target accounts, personalize outreach, handle objections and find patterns towards a quick close. If the pipeline runs dry, they need to diagnose issues and determine whether it’s their own missteps, or an issue with another team’s outputs, such as product or engineering.

Current agents are very good at completing tasks, especially when the output can be verified immediately. But they are surprisingly bad at making informed decisions over long time horizons.

Longer Context Is not Enough

Most agents implement some variation of ReAct: the model reasons about what to do, takes an action, observes how the world responds, and reasons again. For a while, practitioners believed the path toward long-horizon agents looked straightforward:

model + tools + loop + infinite context = autonomous agent

The corollary was that if the context window became large enough, the agent could keep every prior thought, action, and result in view, loop forever, and learn in context.

This is dead wrong. More time = more mistakes. Recent studies find a 30% accuracy drop when long-context models are asked to retain information across sustained interactions. Over longer horizons, agents also suffer from what researchers call "behavioral state decay": requirements, prior attempts, diagnoses and open sub-goals begin as priorities, but at some point stop influencing the agent's decisions.

Long context gives an agent a larger transcript. It does not decide what remains important, which beliefs should be revised, or whether the current plan still serves the original goal. Essentially, long context just offers an agent more time to make mistakes and lose the plot.

This forgetfulness is expensive. Repeatedly looping agent jobs also incurs higher token costs. We’re seeing increasing concerns from orgs like Uber and ServiceNow that have burned through annual token budgets in a matter of months as the “free lunch” effect from tokenmaxxing has come to an end. Token costs have gotten so bad, that there’s a growing trend of inference theft leading to eye-opening losses via public gateways. What organizations need to make agents live up to the hype is the ability to build and run them over long-horizon jobs without breaking the bank.

The Five Missing Systems

Long-horizon agency isn’t as simple as a model running in a loop for a long time. We’re missing critical systems that provide continuity of intelligent behavior across changing evidence, context boundaries, process failures and learning cycles. We’re just not there yet, but I’m hopeful.

We believe continuity depends on five systems:

Organizational memory. Most agent memory is scoped to a single session. Long-horizon agents need useful knowledge to survive beyond that session, evolve as facts change and become available to other agents in the organization. Examples of companies working here include Mem0, Mubit, and Antfly.

Goal orchestration. Distant outcomes do not provide enough feedback to guide each decision. Agents need to break them into intermediate goals that make progress measurable without losing the relationship between a local reward and the original objective. Examples of companies working here include LangChain, CrewAI, and Paperclip.

Durable execution. Long-running work must survive crashes, expired credentials, delayed approvals and irreversible interactions with the outside world. The execution layer must preserve state and know whether a failed operation should be retried, reversed, escalated or stopped. Examples of companies working here include Temporal and Diagrid.

Learning. A transcript records what happened; learning changes what happens next. Agents need a way to turn outcomes into better playbooks, instructions, memory and, potentially, model behavior without allowing every noisy experience to rewrite the system. Examples of companies working here include: Letta and Modiqo.

Toolchain Ecosystem. Agents need additional tooling to extend their capabilities into real-world use cases where they can genuinely move the needle. Among others, a competitive ecosystem of agentic tooling that offers better harnesses, secure sandbox environments and verification, comprehensive monitoring, the ability to support bespoke, custom builds for individual use cases, and large-scale governance and security need to emerge. Examples of companies working here include: Galileo, AgentOps, Halluminate, Bronto, and Strix.

Delineations between these system categories will blur. Memory, for example, may eventually absorb some of what we now call learning. Some capabilities may become native to models. Others will remain properties of the harness, runtime, or organization around the model. The primitives powering these systems are still being designed and implemented.

Got Opinions or Want to Hear More?

If you’re interested in exploring these further or you’re already building in one or more of these categories, join us on Sept 24 in SF alongside Glean, Hex, and Baseten at DevGuild: Long-Horizon Stack. In future posts, I’ll explore each of these systems and how orgs can transform agents from order takers to outcome owners.