Skip to content
OpenCorvus

Where long-horizon work breaks

The three ways long-running agent work fails — stopping short, unusable results, and never improving — and the mechanism in OpenCorvus that answers each.

Long-horizon capability is a property of the whole model-and-harness system, not of the model. A capable model inside a harness that loses task state still stops short; a harness that never checks what came back still hands you something you cannot use.

Three failures account for most of it. This page names each one and the mechanism that answers it, with the part of the system that owns it.

1 · It stops short

The run ends before the work does — a step is skipped, a process dies, or a Task reaches a terminal state with the goal only half met.

Scope is written down before execution. The Requirements stage emits REQ-N entries, each carrying its own acceptance condition, its explicit non-goals, and references to the evidence it rests on. Decomposition is the Architect’s, not a side effect of whichever agent went first.

Order is a declared dependency, not a convention. An Expert Squad’s workflow declares which agents run and what each depends on. A Task pins one workflow for its lifetime.

A lost process is recovered, not restarted. Physical ownership is an append-only lease. When a process disappears, the reconciler terminalizes the exact abandoned assistant Turn at the lease’s expiry — a deterministic timestamp, so two concurrent reconcilers converge on the same completion — and only then acquires a successor. Every accepted input is reduced through one total order in which each state has a name; nothing is silently dropped, and an exhausted budget is an explicit fence rather than a quiet stop.

Terminal is not final. A Task that reached completed, failed, or cancelled reopens when you send it a message. That opens a fresh execution occurrence at the next epoch; the previous occurrence stays intact as an immutable fact. There is no separate retry or replan control to find, because a state whose only exit is special vocabulary is a state you cannot leave with an ordinary action.

2 · The result is not dependably usable

The run finishes and reports success, but what you get is a summary you cannot check, a file that does not open, or a number nobody re-derived.

Handoffs are typed artifacts, not prose. A worker publishes an Artifact with provenance and an exact locator; the next worker reads that locator. The causal read boundary is the producing Provider step, so only completed prior-step outputs are visible — a sibling running in parallel cannot leak an unfinished result forward.

The host records facts separately from the agent’s account of them. File changes and command results are host observations. They exist whether or not an agent mentioned them, which is what makes an agent’s summary checkable rather than authoritative.

Verification runs as named stages. Fact-checking, integrity review, and visual QA are steps in the workflow with their own agents and their own outputs — not a line in a prompt asking the writer to double-check itself. Where a deliverable is a qualified file — today that is the editable presentation profile — the Work Artifact path renders and inspects it, and the validation receipt is what makes it delivered.

3 · It never gets better

The tenth run repeats the mistake of the first, because nothing about what you corrected survived the conversation it happened in.

A squad can be revised from what you told it. State a durable preference — one that would apply again to the next task of this kind — and the host copies the exact installed revision, applies the edits, validates the result as a runnable package, and stages it against the preference you stated. It installs when you accept it. See How squads evolve.

Or from measurement. Evolution Lab freezes the target revision, cases, scorers, environment, arm order, budget, and mutation surface before any candidate exists, runs the arms, and produces an integrity review and a comparison recommendation.

Neither path installs anything on its own. Promotion, restoration, and feedback revision each require a real operator confirmation message bound to that exact Project, Task, and root Session.

The real boundary

Long-horizon is not unlimited execution and not permanent unattended operation.

Work continues only while your OpenCorvus runtime is online. Output quality still depends on the model you selected, the sources it can reach, the capabilities you installed, and the evidence available to the run. External effects remain bound by permission rules.