Towards long-horizon agents

How much an AI system remembers between conversations depends on the setup. In a plain chat, most of what a conversation establishes is gone at the end, with the next one starting from roughly the same baseline. Memory features soften this, bringing a synthesis of preferences and recurring topics from one chat into the next. Agentic AI, which completes entire tasks rather than only answering questions, goes further. With an agentic tool such as Claude Code, the person manages the sessions and what passes between them. On a platform such as Claude Cowork this happens in the background. A session is compacted mid-task rather than cut off, and long tasks continue in the cloud without conversation timeouts or perceived context limits. All of this continuity is real, and all of it is automatic. The platform decides what survives.

Continuity of a different kind can be built on top of these platforms. Before the summer break of 2026 we began working this way internally at Supernomial, with long-horizon agents. A long-horizon agent picks up each session where the previous one ended. In practice this behaves like working with the same colleague, day after day. The agent accumulates a persistent working memory and, with it, something close to a persistent personality.

This article sets out why long-horizon agents matter for transfer pricing teams, what they are, and how the mechanics work.

Where the ceiling on delegation sits

All of this continuity matters for two plain business questions, how much of a team’s workload can be handed to agents, and whether the agent truly feels like a continuously learning colleague working alongside the team. A task can only be fully handed over when the agent already knows what the task builds on. What the agent does not remember, someone on the team would have to supply again.

Existing general-purpose agentic AI adds continuity to some degree, but it is managed by the platform rather than by the work. When a long thread approaches the model’s context limit, the platform compacts it, replaces the older part with a summary, and drops the details. The summary keeps what a general compaction judges important, not necessarily what information will be needed the next day or season, and the details it drops are not recoverable. Memory, in turn, synthesizes preferences and recurring themes rather than recording decisions with their reasons. Newer agent platforms go further and give an agent a durable memory store, with automatic tidying of what accumulates there. The information may then be organized, but not necessarily to the degree or in the manner the team wants, because the organizing follows the platform’s general scheme rather than the work’s. The person directing the work therefore remains the memory of the system, keeping track across threads and longer periods of what was decided, what was tried, which approach failed and why. That puts a ceiling on delegation. Work that fits inside one thread and survives its compactions can be handed over. Work that stretches over longer periods and across many separate threads mostly cannot, because the accumulated history would need to be re-supplied at every step.

Transfer pricing work sits mostly above that ceiling. A transfer pricing analysis runs for months, and the hundreds of tasks in it are connected by intermediate decisions. Someone tracks which entity’s numbers arrived late and what workaround was agreed, how a method choice was justified, and what the tax authority questioned in last year’s audit. An agent can draft any single piece of this competently. Succeeding as a competent assistant over the life cycle of an engagement would mean trusting an automatic compaction with decisions it cannot know matter.

What a long-horizon agent is

A long-horizon agent is an agent whose work builds on what it has accumulated across earlier sessions, where a session means one continuous working conversation or one self-run stretch of work between agents. A conventional setup treats each session as complete in itself, and when a session runs long, the platform decides what is kept and what is dropped. A useful long-horizon agent for domain knowledge work knows its domain, treats each session as one shift in a longer piece of work, and keeps its own record of the whole.

The construction is deliberate because no layer underneath does the full job. The AI model, the engine doing the actual reasoning, considers a limited amount of text at once and retains nothing from one run to the next. The platform around the model adds the automatic continuity described above, and increasingly a place where an agent can keep durable notes. The open question on any platform is what gets written there, when, and in what structure. Long-horizon agents answer it with a shift handoff. The worker going home leaves notes, and the worker coming in reads them before starting. The agent does the same for itself, and its handoff is composed deliberately rather than compacted automatically.

The agent’s record extends beyond task notes. It includes the goals the agent is pursuing, the commitments it has made, and the outcomes of earlier decisions. That breadth is why it behaves like the same colleague.

A well designed memory system

The difference is easiest to read as a curve. In a session-bound setup, one where each conversation stands alone, what the agent knows about the work climbs during a conversation and falls back when the conversation ends. The floor it falls back to is the standing context the platform loads into every conversation, the saved instructions and the project documents. Adding to that standing context raises the floor and leaves the shape unchanged. Platforms with persistent threads, compaction, and memory soften this shape. They stretch the sessions and raise the floor, and the curve still steps down at every compaction. In a long-horizon setup, the right information is written down before it can be lost, so the next session starts near where the previous one stopped and the curve climbs across sessions.

The left chart of Figure 1 is where the familiar improvements act. Better prompting, richer context, and memory features lift its floor. The right chart shows the structural change. The agent keeps its own record, and each session starts from what the previous one knew.

The increase in knowledge displayed in the right chart comes from a disciplined working routine, shown in Figure 2. The persistent memory is a structured system of records rather than a single document, reaching from current priorities and progress reporting to the underlying data model. When a session starts, the agent grounds itself in that system and picks up the current goal and the next step. While working in incremental steps, it writes down decisions and facts at the moment they happen. And before a session ends, it brings its records up to date for its future self, so the next session starts where the previous one stopped.

Investing in a durable memory system moves the ceiling on what teams can delegate to agents, one recorded decision at a time.