Artifact Commons20 Reusable Agent Baselines
Independent building blocks for agent work. Each download contains a specification, executable Python, an input schema, examples, named verification cases, file digests, and an MIT license.
| Artifact | Use it to |
|---|---|
| Task Contract Checker | Turn acceptance criteria and evidence references into an explicit completion report. |
| Evidence Ledger | Keep claims, source references, support, and contradictions linked without hiding uncertainty. |
| Resumable Handoff Capsule | Package verified work, unresolved work, and the next action so another agent can resume. |
| Action Precondition Checker | Compare a proposed write with current state and exact approval references before execution. |
| Scoped Idempotency Key | Derive a repeatable operation key from tenant, operation, intent ID, and payload. |
| Explicit Retry Planner | Calculate the next retry time from a caller policy without silently retrying writes. |
| Capability and Permission Matcher | Select eligible tools by declared capabilities and permissions, with reasons for rejection. |
| Structural JSON Change Map | Produce a deterministic, pointer-addressed explanation of what changed between JSON values. |
| Dependency Order Planner | Turn prerequisite lists into deterministic execution layers and expose blocked dependencies. |
| Constraint First Decision Table | Reject options that break stated constraints, then compare eligible options with explicit weights. |
| Budgeted Context Selector | Fit reusable context into an explicit budget while preserving every required item. |
| Evidence Freshness Check | Separate fresh, stale, and future-dated observations using explicit per-item freshness rules. |
| Untrusted Content Envelope | Keep retrieved text and its source separate from instructions, with stable content digests. |
| Source Span Extraction Audit | Check that extracted fields point to exact source text instead of unsupported paraphrases. |
| Explicit Span Redactor | Remove declared sensitive spans and record where replacements moved in the output. |
| Dimension Checked Quantity Converter | Convert physical quantities with exact rational arithmetic and reject incompatible dimensions. |
| Time Zone Availability Intersection | Find shared availability across explicit time windows and render it in participant time zones. |
| Lossless CSV Intake Profile | Inspect CSV structure and formula-like cells while keeping identifiers and values as text. |
| Comparable Evaluation Scorecard | Compare baseline and candidate results case by case, keeping regressions and missing coverage visible. |
| Portable Text Integrity Manifest | Build a sorted SHA-256 inventory for text bundles without filesystem or network access. |
Run a baseline
Open an artifact, download its ZIP, and verify the SHA-256 in the document. Unpack it and enter the package directory.
python3 baseline.py < example.input.json
python3 -m unittest -v
Python 3.10 or newer. The packages use the standard library. Time zone rendering also needs the host IANA time zone database. Each package accepts JSON on standard input and returns a JSON result; it can also be imported through run(input_dict).
Selection
These baselines address recurring work where a small explicit contract can prevent repeated mistakes: losing evidence, resuming with stale context, repeating side effects, mixing data types, or reporting success without comparable results. They are original reference implementations, not a ranking of existing products.
Policies stay with the caller. Budgets, retries, freshness rules, permissions, and decision weights are explicit inputs. Each specification names the checks the package performs and the judgments it cannot make.
Verification
All 20 archives passed their packaged checks before publication, including 64 named behavior cases. Live verification checks the published specification, HTML page, exact revision, download digest, and directory discovery. These checks establish the stated baseline behavior; they do not establish fitness for every deployment.