It gets the same ten questions right. Is it ready?

Keep some realistic questions separate from practice. Getting better at the practice set is useful, but it is not yet a reason to rely on the AI with new messages.

A colleague tests both versions on 30 fresh questions with answers agreed in advance. The new instructions sort more correctly, including urgent questions that need staff. Staff have time to review suggestions, but no one has tried the helper alongside a normal working day.

What is a useful next step now?

• Let the AI send and route all messages without staff
  Why this choice?

  The fresh test supports trying the helper, not handing over the inbox. A normal day may contain questions the test missed. Keep staff responsible for replies and routing while you learn how it behaves.

• Keep editing until it gets every fresh test question right
  Why this choice?

  That would turn the separate test into another practice set. Save this result, try the unchanged version beside staff, and use a new untouched test for later changes.

• Try its suggestions beside staff without sending or moving messages
  Why this choice?

  Keep the tested instructions unchanged. In the centre's approved workspace, let staff compare suggestions with their own decisions and note mistakes and review time. Keep the old version available. A small observed trial can show whether the improvement helps during real work.

Practise on one set of questions; check progress on another.

How could you tell the difference between learning the task and learning the practice answers?

Held-out evaluation uses questions kept out of improvement work. Shadow mode lets a candidate make suggestions alongside the current process without acting on them. Here, only the instructions changed; the underlying AI model did not.

Separate prompt search, runtime changes and weight training, then decide what independent improvement evidence warrants.

Evaluation and adaptation · 10 minutes

The problem

Synthetic practice scenario: all quantities and evaluation outcomes are illustrative, not reported research. A team wants a support-routing agent to improve from failures, but its current optimizer has repeatedly seen the same evaluation tasks.

Your initial decision

What should the team do before declaring improvement?

• The illustrative optimizer has selected prompts against the same 80 cases used in the improvement chart.
• The current model weights are frozen; proposed changes affect tool-error handling and context assembly.
• The optimizer must not edit graders, hidden tests or the production permission policy.
• Customer transcripts are not approved for training or synthetic paraphrasing; new tasks must use permitted specifications.

Compare the alternatives

• Repair the harness and establish a clean holdout

  A manual harness correction plus independently authored tests is slower than accepting the search chart, but creates interpretable evidence.

• Continue bounded search for candidate artifacts only

  Search may find useful prompt or runtime candidates, but reused tasks cannot support a release claim without separate evaluation.

• Train a routing model on expert-labeled synthetic tasks

  Supervised tuning on permitted, independently labeled specifications could improve stable routing behavior, but adds dataset and checkpoint maintenance before the runtime defect is isolated.

Mechanism, trade-offs, and evidence

• Name the changed object: this proposal changes the harness around frozen weights, not the model weights or necessarily the improvement procedure.
• An RSI survey is a taxonomy and research roadmap, not proof that each local iteration compounds or generalizes.
• The existing Meta-Harness source distinguishes held-out classification and math results from coding results that reuse search tasks; preserve that distinction here.
• Keep training, selection and final evaluation separate, and do not assume synthetic rewriting removes privacy or contamination concerns.

Reconsider after a change

An independent evaluator has tested a frozen harness candidate on a previously unseen task set with protected graders and a documented data origin.

What rollout, if any, is justified by the new evidence?

• In the illustrative 200-case holdout, the baseline routes 160 correctly and the candidate routes 178 correctly; both use the same frozen model and call budget.
• Reviewers independently adjudicated disagreements and checked critical routing and refusal slices; finite coverage still leaves residual risk.
• The synthetic latency target is met, and rollback restores the previous harness artifact without changing model weights.
• The owner can monitor a read-only shadow rollout before approving a narrow live canary; no automated permission changes are allowed.

Compare the alternatives

• Start an independently evaluated supervised tuning study

  A small routing checkpoint might eventually reduce inference costs, but the held-out gain already came from a harness fix and supplies no evidence that tuning adds value.

• Promote the frozen harness to monitored shadow evaluation

  Run it beside the baseline without taking customer-facing actions, then consider an owner-approved canary if live-distribution evidence holds.

• Keep the manual repair and expand targeted evaluation

  Retaining the baseline repair limits change while experts add missing risk cases, but delays the independently measured benefit of the selected harness.

Mechanism, trade-offs, and evidence

• The decision changes because independent evaluation now exists and isolates the harness change; the numerical results are fictional practice facts, not claims about the cited papers.
• The aggregate gain is not enough alone: inspect critical slices, resource use and reviewer disagreements before interpreting it as useful.
• Shadow mode checks transfer to current traffic without letting the candidate act; live execution still requires explicit owner approval.
• After promotion, preserve this holdout as a recorded final evaluation and create fresh protected evidence for later candidate selection.

Related technical concepts

Follow-up lessons retain their existing sign-in and plan requirements.

• self improving agents
• agent harness engineering
• prompt optimization patterns
• fine tuning
• agent learning
• synthetic evaluation data
• agent testing benchmarks
• experimentation continuous improvement
• agent evaluation methodologies

Technical sources

• Demystifying evals for AI agents — Anthropic Engineering

  Explains tasks, graders, repeated trials and outcome evaluation for agents; use task-specific tests rather than importing reported scores.

• The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement — Duan et al. / arXiv

  Survey and roadmap separating persistent adaptation from changes to the improver; preliminary cases do not demonstrate inevitable recursive progress.

• Meta-Harness: End-to-End Optimization of Model Harnesses — Lee et al. / arXiv

  Searches executable harnesses around frozen weights; held-out classification and math differ from coding results that reuse search tasks.

• AI Risk Management Framework — NIST

  Organizes voluntary risk management around governing, mapping, measuring and managing; it is neither certification nor a return-on-investment study.

Explore the technical decision