Sixty harness primitives converge on one durable loop: log everything, gate before acting, verify after, stop on budget, resume from the log

Asked (summary)

Research public documentation, engineering posts, open-source repositories, papers, and release notes for Claude Code, OpenHands, Cursor, Devin, Aider, and other long-horizon AI agents. Return one row per reusable agent-harness primitive — with lifecycle stage, trigger, persisted state, event-log representation, tool interface, permission behavior, stop condition, verification, retry/restart behavior, public evidence URL and strength, and applicability to a persistent social-media agent — separating directly observed behavior from inference and excluding unsupported marketing claims.

The evidence-screened set holds 60 reusable harness primitives drawn from papers, official docs, open-source repositories, and engineering posts across 14 independent source hosts; 48 rows are corroborated on multiple hosts and no single evidence URL backs more than 3 of the 60 rows. This is a screened research sample of publicly documented behavior, not exhaustive coverage of every private product feature.

Where each primitive sits in the harness lifecycle

Hover a square for the primitive and its products; click a square or cell to filter the table below

Durable state dominates: 12 of the 60 primitives are append-only or event-sourced logs whose replay gives pause, resume, fork, and audit "for free" — OpenHands models the agent as a pure function from event history to next event (dev.to).

Thirteen primitives are pre-action gates — risk-scored confirmation pauses, sandboxes, secret registries — such as OpenHands' SecurityAnalyzer that holds the agent in a WAITING_FOR_CONFIRMATION state until a human approves (arxiv.org).

Stopping is contractual, not emergent: Claude Code's loop ends only on no-tool-call responses, max_turns, max_budget_usd, or error, and its Stop hook can refuse a halt and re-activate the agent via exit code 2 (github.com).

Magentic-One's per-round progress ledger detects stalls and forces a bounded reset-and-replan — the clearest published recipe for keeping a long-horizon loop from spinning in place (learn.microsoft.com).

Reading the rows: the green "Directly observed" text restates only what the cited source documents or implements. The orange "Inference:" text is analyst reasoning about transfer to a persistent social-media agent — it is not a product claim. A dash marks a field the source does not publicly state; no value was invented.

Primitive	Products	Phase · stage	Family	Evidence	Source

Method: 60 normalized reusable agent-harness primitives from an evidence-screened research corpus of papers, official documentation, open-source repositories, engineering posts, and release notes across 14 source hosts; one row per primitive, each with its own public evidence URL (no URL backs more than 3 rows). Evidence strength distinguishes multi-host corroboration (48 rows) from multiple technical records on one host (12 rows); supporting-record and supporting-host counts come precomputed with the rows. Lifecycle phases and primitive families are analyst groupings of the source-stated lifecycle stages for the matrix; exact stage text appears in the table. Fields the sources do not publicly state are shown as dashes, not filled in. Long observed-behavior and inference texts are truncated in the corpus at capture; nothing else was cut.

This report was generated automatically by Keenable SELECT at a user's request, from publicly available web sources linked herein. Keenable does not review, verify, or endorse its contents and makes no representation as to accuracy, completeness, or timeliness; AI-based extraction may contain errors. Nothing in this report is investment, legal, financial, or other professional advice. All trademarks and referenced content remain the property of their respective owners; no affiliation or endorsement is implied. To report an error, rights concern, or request removal: legal@keenable.ai.