How AgentTape works.

1. Discovery — how agents enter the index

Nothing on AgentTape was added by hand. A discovery service sweeps a fixed list of public sources on its own schedule and opens a candidate row for anything matching an AI-agent or foundation-model pattern. A second pass scores each candidate and either admits it, rejects it, or leaves it pending for weekly review.

Sources swept: GitHub search (repos matching agent-frame patterns), Hugging Face trending, OpenRouter models catalogue (every published foundation model), MCP registries, npm and PyPI, arXiv, and the Hacker News firehose. Each source ingests on its own cadence — none of them gate on the others.

Promotion scores each candidate on five axes: LLM dependency, agent-vocabulary match, popularity floor, maintenance, and packaged distribution. Four small substance bonuses (each capped at 0.05) tip borderline cases — declared topics, description length, multi-ecosystem packaging, recent commit activity. Archived or disabled GitHub repos are hard-rejected.

2. Refresh tiers — how often signals update

Signals refresh on two cadences. The split balances ticker freshness against upstream rate limits — anything hitting a benchmark site or a slow JSON API runs once a day; everything else runs hourly.

3. The AgentScore

One number, 0-100, computed as a weighted sum of pillar scores. Applications have four pillars; foundation models have five (Efficiency adds cost + speed, which doesn't apply to apps that run on user hardware).

Weights differ by entity kind because the question "what makes a coding tool good" is not the question "what makes a foundation model good". A 70 for an app and a 70 for an FM are not directly comparable — use the per-kind boards (Models, Sectors) when you want a fair comparison.

A pillar with no signals contributes zero to the headline (no redistribution, no re-normalisation). Less data is a lower ceiling, not a re-weighted average — a 3-pillar FM caps at 80, a 4-pillar one at 90, only a fully-rated model can reach 100. That keeps coverage honest. Models board lets you click any column header to re-sort by that single pillar.

4. The pillars

What each pillar answers, and the signals that drive it for each entity kind. Signal lists below are the source of truth — they're co-located with the source code in apps/scoring/compute.py and the page renders straight from that list.

5. Scoring formulas

Three families of formula, applied per signal kind. Each produces a 0-100 contribution; the pillar score is the arithmetic mean of available contributions.

Counts (most signals)

scaled(v, anchor) = min(100, 50 × log₁₀(v + 1) / log₁₀(anchor + 1))

A log curve so a project with 1,000 stars scores 50 and one with 100,000 stars doesn't get 100× the credit. Each anchor is the value at which the signal scores exactly 50 — chosen so the median agent in each population lands near the middle of the scale. Anchor table lives in source at apps/scoring/compute.py:ANCHORS.

Benchmarks (Quality pillar)

Mean percentile rank across the agent's benchmark coverage. For each benchmark the agent has been scored on, its score is ranked against every other agent on that benchmark and converted to a 0-100 percentile. The pillar score is the mean of those percentiles.

Percentile rank is coverage-robust: a model with 5 benchmarks all at the 95th percentile beats a model with 10 benchmarks averaging the 70th. It's also head-to-head consistent — if A strictly beats B on every shared benchmark, A's mean percentile is ≥ B's. The previous formula (mean of normalised scores) failed this: a model could outrank a strictly-better competitor just by having extra easy benchmarks pulling its mean up.

A coverage floor of three distinct benchmarks is required to rate a model on Quality — below it the model stays Unrated.

Momentum (7-day rate of change)

roc_7d     = (now − then) / max(then, 1)
scaled_roc = clamp(50 + 50 × roc_7d, 0, 100)

0% growth → 50, +100% → 100, −50% → 0. A signal first seen inside the 7-day window (no "then" reading) gets a 60 — a small "newly arrived" bias, not the punitive 0 a missing baseline would otherwise imply.

Special cases

A handful of signals don't fit the count log-curve. MCP registry listed is binary (0 → 0, 1 → 75 — a hand-picked credibility bonus). HF trending rank and GitHub first-response hours are inverted: lower input = higher score, with the same log shape mirrored. Cost (blended price) is also inverted — cheaper scores higher, anchored at $5/M tokens.

6. Why some agents are Unrated

A pillar with no signals on file is Unrated rather than zero. The pillar's card on an agent page reads "Unrated" (no data), while at the headline level that pillar contributes zero (because the headline is a weighted sum and you can't add an unknown to a sum). The two read like a contradiction; they describe different layers.

Quality has the additional coverage floor — fewer than three benchmarks means Unrated rather than a noisy single-source score. New flagship releases often sit Unrated on Quality for a few days until enough leaderboards pick them up.

7. Manipulation resistance

Three patterns trigger automatic flags. Flagged signals are excluded from that day's score; the agent's record carries the reason. The manipulation_resistance confidence on every score envelope reflects how clean the inputs were.

• star_spike_no_contrib_diversity — a 10× star jump in 24 hours from very few distinct contributors. Excludes github_stars for that tick.
• hf_surge_no_github — a Hugging Face download surge with no accompanying GitHub activity. Excludes hf_downloads_30d.
• coordinated_hn_posting — a burst of HN mentions with low account-age diversity. Excludes hn_mentions_7d and hn_points_7d.

8. Indexes

Six indexes at launch — each with eligibility rules published in code, equal-weight v1, rebalanced Mondays at 03:00 UTC. Every diff is logged with a short narrative.

• TAPE-100 — top 100 across both kinds.
• FM-50 — top 50 foundation models.
• CODE-25 — top 25 coding agents (capability:code-generation, application).
• WEB-25 — top 25 browser agents (capability:browsing, application).
• OSS-50 — top 50 open-source applications.
• MCP-25 — top 25 MCP servers (deployment:mcp-server).

9. Show your work

Every agent page exposes its raw signals as a downloadable CSV. Every index page links its rebalance log. Methodology changes are versioned in the repository; corrections welcome via Issues or Discussions.