Offline Consolidation Cycle
A scheduled, offline pass in which an agent mines its own
past sessions for durable lessons, re-runs recurring tasks to test
candidate improvements, and promotes only what survives a gate into its
live skills and memory. The work happens when the agent is not
serving requests — hence SkillOpt's frank borrowing of the sleep
metaphor, complete with
harvest → mine → replay → consolidate,
dream.py (rollouts on imagined/perturbed variants of past
tasks), experience replay, and long-term memory. The lesson the operator
learns from real use is deposited in the artifact, not in the operator's
head.
The structural claim worth extracting is about where the improvement work lives. Three positions exist in the graph. Improvement can happen in the request (reflection, evaluator-optimizer — pay per request, forever). It can happen between sessions, by a human (Self-Improving System's periodic loop — pay in attention, the scarcest thing). Or it can happen offline, mechanically, on a schedule — pay in compute at night, and pay nothing at serve time. Only the third scales in the dimension that actually binds: the operator's attention. That is the argument for consolidation cycles, and it is a good one; the price is that whatever the gate does not measure, nobody is watching.
Two details distinguish this from ordinary Agent Rituals scheduling. The cycle's
input is the agent's own trajectory history — real past
sessions, harvested from Claude Code / Codex / Copilot transcripts — so
the training distribution is the operator's actual work, which is the Self-Improving System instinct
made into a data pipeline. And the cycle is budgeted
(budget.py) and staged
(staging.py), meaning the run has a spend ceiling and
promotion is a separate act from generation.
Claims
- Move improvement work off the request path and onto a schedule: mine past sessions, replay recurring tasks, promote what passes a gate. best practice — context: an agent used repeatedly on a recurring distribution of tasks, where session transcripts are retained and idle compute is cheaper than operator attention. It is not the right shape for one-off work or for a task distribution that shifts faster than the cycle runs, since a nightly cycle optimizes for yesterday.
- An agent's own past sessions are its best training distribution, because they are drawn from the work it will actually be asked to do again. principle — durable; it is why Error Analysis on real traces beats imagined failure modes, restated for skills instead of evals.
- Consolidation must be staged and gated, not applied — generation of a candidate skill and promotion of it into the live directory are separate acts. best practice — context: an unattended nightly loop with write access to the artifacts the agent runs on tomorrow. See Validation-Gated Update; without the split, "self-evolving" means "unreviewed writes to production."
- The scarce resource in a compounding system is operator attention, not compute — so the design question is which improvement work can be moved off it. principle — durable. Inference-time reflection spends money per request; a human improvement loop spends attention per cycle; an offline gated cycle spends compute once. Attention is the one that does not scale.
- The cycle is budgeted and multi-objective, and includes "dream" rollouts on variants of past tasks alongside literal replay. observation — the repo's stated v0.2.0 mechanism; whether dreamed variants improve generalization over plain replay is an empirical claim the paper, not the README, would settle.
Check-worthy source claims (attributed, not adjudicated — a later grounding pass can verify):
- SkillOpt-Sleep ships as a
skillopt-sleepCLI in v0.2.0 (2026-07-02) with harvest backends and plugin shells for Claude Code, Codex, Copilot, Devin, and OpenClaw. observation
Related
- Validation-Gated Update — the promotion rule; the cycle is only safe because the gate is.
- Text-Space Optimization — the same optimizer, pointed at a live agent's history rather than a benchmark task set.
- Self-Improving System — the human-driven periodic loop this mechanizes; the disagreement is over whether the human's sign-off can be replaced by a gate.
- Agent Rituals — scheduled routines are the harness-level scaffolding; a consolidation cycle is the heaviest such routine, and the only one that rewrites the skills the rituals invoke.
- Layered Agent Memory — what the cycle writes to: long-term memory and skills, the slow layers, updated on a slow clock.
- Error Analysis — mining past sessions for what went wrong is error analysis with the human removed from the loop; the same "look at your data" principle, automated.
- Decision Log — the natural audit surface for an unattended loop: what did it promote last night, and on what evidence.
- Agent Supervision — the open question a nightly self-evolving cycle raises and does not close: who reads the changelog.
- Skill-Driven Loop Development — skills as the unit of improvement, here improved by machine.
- Distillate: microsoft/SkillOpt — training skills like weights, without touching weights