Skip to content
Article reader Listen + reading controls
LISTEN + READ YOUR WAY

Article reader

Preparing the reader…

0:00 0:00
Reading settings
Text size
100%

Long-horizon agents need short feedback loops

Anthropic's September 29 release of Claude Sonnet 4.5 highlights stronger coding, computer use, and sustained work on complex tasks. Alongside the model come checkpoints for Claude Code, a memory tool and context editing for longer agent runs, and an agent software development kit (SDK).

The pairing is instructive. Longer autonomy arrives with better ways to see, constrain, and reverse the work.

An artificial intelligence (AI) agent that can maintain focus for hours can take on work otherwise too fragmented for automation. It can explore a codebase, form a plan, use tools, recover from failures, and carry changes across many components.

Each additional step also creates another opportunity for a small error to alter the remaining path.

Horizon length is not the same as authority

An agent may need a long time to complete a task without needing broad permission to act. Those dimensions should be designed separately.

A useful authority envelope defines which repositories, tools, data, environments, and actions are available. It distinguishes reversible work from consequential changes. It specifies when the agent can proceed, when it must request review, and which conditions stop the run automatically.

For example, an agent might work independently in a branch and test environment for hours while remaining unable to merge, deploy, change access controls, or reach sensitive production data.

That is meaningful autonomy within a jurisdiction.

Checkpoints create recoverable work

A checkpoint is more than a convenience feature. It changes the economics of oversight. When state can be inspected and restored, people can allow exploration without making every intermediate action permanent.

Effective checkpoints should coincide with meaningful transitions: after planning, before a broad refactor, after tests establish a new baseline, before an external side effect, and at the point of handoff. The system should summarize the changes, the evidence it produces, and what remains uncertain.

David Woods's work on resilience emphasizes the capacity to adapt when conditions challenge the plan. In agentic work, inexpensive rollback, visible state, and alternative paths are practical sources of that capacity.

Feedback must arrive before the error compounds

Human review only helps when it occurs while intervention is still useful. A final inspection after thousands of poorly directed changes may be technically possible and operationally hopeless.

Long tasks need short feedback loops through tests, static analysis, resource limits, policy checks, intermediate previews, and selective human decisions. The frequency should follow risk. A low-impact research task can run broadly; a change to an authentication path needs tighter gates.

Evaluate the system over a trajectory

Single-answer benchmarks reveal little about an agent's ability to recover, recognize uncertainty, and preserve intent over time. Evaluations should include interrupted tools, ambiguous requirements, changing conditions, misleading intermediate results, and opportunities to ask for help.

The desired behavior is not blind persistence. It is progress with calibrated escalation.

Long-horizon agents will be valuable because they can remain with difficult work. They will be trustworthy when the surrounding system keeps that work observable, bounded, and recoverable. The longer the horizon, the shorter the feedback loop should become.

Sources and research trail

READER-NEUTRAL SUBSCRIPTION

Follow Field Notes via RSS.

Copy this address into the RSS reader you already use. New notes will appear there automatically—no account, email address, or tracking required.