Seven shapes, 35 jobs built from them

Find the recurring mistakes in a team's agent transcripts and fold them into a short list of themes with examples.

Checked by recall of planted defects

Answer a hard decision question from a large pack of sources, every claim quoting its source.

Checked by blind out-of-family judges, receipt audit

Map a large monorepo into a navigable guide: what each area does, who calls it, what is fragile.

Checked by fresh agents do tasks faster with it than without

Compare hundreds of contracts against a standard and flag every deviation with the clause quoted.

Checked by hand-checked sample of flags

Find places where an organisation's docs, prompts and memories disagree with each other.

Checked by planted contradictions, owner accept rate

Read every ticket, review and call note from a quarter and rank the themes by revenue at stake.

Checked by sampled theme check against raw tickets

Read a full data room and produce the risk memo with page-level citations.

Checked by expert spot check of cited risks

Turn a year of meeting transcripts into a log of decisions, owners and reversals.

Checked by participants confirm a sample

Distil a shelf of technical books into an index agents can consult during design work.

Checked by held-out design tasks with and without it

Turn a long regulation into the concrete obligations a product must meet, mapped to features.

Checked by compliance reviewer sample

Read the patents in a field and cluster them by claim, holder and expiry.

Checked by hand-checked cluster sample

Read every call in a sector for one quarter and extract guidance changes with quotes.

Checked by quotes match transcripts verbatim

Reconcile out-of-date docs with what the code actually does and flag every mismatch.

Checked by mismatches reproduce against the code

Keep a standing one-page brief across many projects that updates only where something changed.

Checked by owners mark overturned items

Screen, read and extract a field's findings into a table where each row quotes its paper.

Checked by receipts per row, sampled hand check

Break an open question into sub-questions, answer each, then attack the answer.

Checked by every claim cites a primary source

Size the market, map competitors and price the risks behind a single recommendation.

Checked by sampled fact check, later outcome

Rebuild the experiment from the paper and say whether the claim holds.

Checked by published numbers within tolerance

Use and document every competitor's product against a feature rubric.

Checked by second reviewer agreement

Rank a backlog of research hypotheses by evidence, cost to test and expected value.

Checked by later experiment outcomes

Estimate the effect of a proposed rule from data, precedents and models, with ranges.

Checked by back-test on past rules

Find whether an invention or idea already exists anywhere in public record.

Checked by expert confirms hits

Search for better constructions on problems with exact verifiers.

Checked by exact verifier, no float tricks

Screen known compounds against a target using literature and public assay data.

Checked by expert review, later assays

Trace a product's suppliers several tiers deep and flag concentration risks.

Checked by sampled supplier verification

Answer a question from scanned archives that no single reader could get through.

Checked by historian spot check

Explore a large survey from many angles and report only findings that survive correction for multiple tests.

Checked by pre-registered held-out split

Count customers and spend segment by segment with sources for every input.

Checked by triangulation against top-down estimates

Recreate a working program from its binary and documentation against hidden tests.

Checked by hidden test suite

Move many repositories to a new library version inside a tamper-proof test runner.

Checked by held-out tests agents never see

Turn a backlog into a dependency graph and build it in parallel with review.

Checked by per-ticket acceptance tests

Rewrite a service in a new language and prove it behaves the same.

Checked by differential fuzzing against the original

Restructure a system and replay recorded production traffic to prove nothing changed.

Checked by differential replay

Raise meaningful coverage across a codebase, module by module.

Checked by mutation score, not line count

Upgrade every outdated dependency across a fleet of services and fix what breaks.

Checked by CI plus canary

Add strict types to an untyped codebase without changing behaviour.

Checked by type checker plus test suite

Fix accessibility issues across every screen of an application.

Checked by automated audit plus screen-reader spot check

Extract every string, translate, and fix layouts that break in long languages.

Checked by native reviewer sample, layout snapshots

Carve a module out of a monolith behind a clean interface.

Checked by contract tests plus replay

Capture hand-built cloud infrastructure as code that plans to zero changes.

Checked by plan shows no diff

Generate and hand-finish client libraries for one API in eight languages.

Checked by shared conformance suite

Plan and execute a large schema change with zero data loss.

Checked by row-level checksums before and after

Build several competing prototypes of one feature and keep the best.

Checked by user test or benchmark

Turn thousands of loose documents into a structured, linked site.

Checked by link checker plus editor sample

Generate and playtest levels to a difficulty curve.

Checked by automated playtest agents

Add a language feature to a compiler against its conformance suite.

Checked by conformance tests

Investigate a live or replayed incident with parallel hypotheses and one commander.

Checked by known root cause on replayed incidents

Chase a slowdown with each hypothesis on its own machine and a registered prediction.

Checked by effect vanishes under the control

Find and fix the root cause of every flaky test in a suite.

Checked by thousand reruns stay green

Split the attack surface and require a working proof of concept for every finding.

Checked by proof runs in a sandbox, planted-bug recall

Reconstruct what happened from terabytes of logs across services.

Checked by timeline matches known events

Find and apply savings across an account without breaking anything.

Checked by bill before and after, error rates flat

Review every alert rule, tune or delete the noisy ones, and keep coverage.

Checked by missed-incident back-test

Execute every runbook in a sandbox and fix the ones that no longer work.

Checked by runbooks pass in staging

Group thousands of crash reports by root cause and fix the top ones.

Checked by crash rate after release

Compare every server's config against its intended state and explain each difference.

Checked by drift report reproduces

Compete on optimisation problems with a hard deadline and scored submissions.

Checked by official scorer on held-out seeds

Make a program faster through many small experiments, each on its own machine.

Checked by benchmark with a noise floor

Improve an agent's prompts and procedure against an eval set.

Checked by held-out eval at stop

Search configurations of a GPU kernel or SQL query for the fastest correct one.

Checked by correctness check plus timing

Find better schedules or routes for a real operations problem.

Checked by cost function on real data

Run a training sweep where agents read curves and decide what to try next.

Checked by held-out validation score

Design and analyse pricing tests across segments.

Checked by pre-registered analysis

Cut a slow build through caching, splitting and dependency cleanup.

Checked by cold and warm build times

Search placements under hard constraints with an exact checker.

Checked by design-rule checker

Attack a set of hard puzzles with verifiable answers.

Checked by exact answer check

Compare models and agent setups on a task set with enough repetitions to trust the result.

Checked by validated grader on planted defects

Mine real tasks from release history where the reference passes and baselines fail.

Checked by reference passes, solver cannot game it

Find failure modes in a model under a written policy, each with a reproducible example.

Checked by reproducible failures, reviewer sample

Check an automated grader against planted good and bad answers before trusting it.

Checked by planted answer recall

Read thousands of agent traces and classify where and why they failed.

Checked by second labeller agreement

Run the same jobs under every swarm shape to learn which shape suits which work.

Checked by matched-dollar verdicts

Find stale, leaked or broken items in an eval suite and replace them.

Checked by items pass reference, fail weak baseline

Assemble evidence for a system's safety argument, claim by claim.

Checked by reviewer accepts each claim

Remove components one at a time and measure what each was worth.

Checked by pre-registered metric

Measure and improve a search or retrieval system query by query.

Checked by labelled relevance sample

Label thousands of items with several labellers and route disagreements to experts.

Checked by agreement with expert sample

Generate, filter and verify data to train a small model on one skill.

Checked by trained student on a held-out eval

Resolve people, companies and places across many sources into one graph.

Checked by hand-checked match sample

Merge hundreds of inconsistent spreadsheets into one clean table.

Checked by row reconciliation totals

Collect a product or parts catalogue from many sites and normalise it.

Checked by sampled field accuracy

Find and redact personal data across a large document set.

Checked by planted PII recall

Translate thousands of articles with terminology kept consistent.

Checked by bilingual reviewer sample

Infer the rules a dataset should satisfy and find every violation.

Checked by owner confirms rules

Extract features from imagery or maps tile by tile.

Checked by ground-truth sample

Build a validated Q&A set from a documentation corpus.

Checked by answers verified against source

Watch a project's benchmarks, investigate regressions, and propose fixes with before and after numbers.

Checked by back-test on history with planted reverts

Watch feeds for changes that matter and send a short digest.

Checked by back-tested recall on a past window

Watch advisories, check reachability in your code, and open patches.

Checked by reachability proof per alert

Track competitors' docs, pricing and releases and explain what changed.

Checked by missed-change back-test

Re-run a shape bench whenever a new model or procedure ships.

Checked by matched-dollar verdicts

Watch pipelines for silent drift and investigate when distributions move.

Checked by planted drift recall

Read new papers daily in a field and keep a living survey up to date.

Checked by expert marks misses

Read the engine's own run records and propose small, reversible procedure edits.

Checked by matched-dollar before and after

Write a long handbook where every number traces to a source.

Checked by fact check on numbers

Write a full course with exercises that are tested and auto-graded.

Checked by exercises pass their own graders

Assemble a long response document against a scoring rubric.

Checked by rubric review by a second team

Write complete reference docs for a large API from its code and tests.

Checked by examples run as written

Write a multi-chapter report where editors enforce terms and numbers across chapters.

Checked by cross-chapter consistency check

Write and test copy variants for many markets.

Checked by native reviewer and live test