How AgentTape works.

No seed list

A seed list would bias the index toward whatever the maintainers already knew about. The point of AgentTape is the opposite — to surface the things you have not heard of yet, the morning they start to matter. The discovery service sweeps GitHub, Hugging Face, MCP registries, npm and PyPI, arXiv, and Hacker News on its own schedule. Admitted agents appear in the index without human intervention.

Three ingestion tiers

Once an agent is admitted, signals refresh on a tiered cadence to balance freshness against rate limits.

• Fast (~5 min) — GitHub stars, HN mention velocity, HF trending rank. Drives the ticker.
• Medium (~1 hour) — forks, contributors, 7-day commits, HF downloads, npm/PyPI counts, Reddit, MCP registry presence.
• Slow (daily) — benchmark scores from Galileo, HAL, AstaBench, LLM-Stats, Steel WebVoyager; arXiv citations.

The AgentScore — the equation

Every agent has a single 0–100 headline backed by four pillars (Adoption 35%, Quality 30%, Momentum 20%, Community 15%). The formula is identical for application agents and foundation models, so any two scores compare directly.

Scaling a raw signal to 0–100

For a count-shaped signal with raw value v and an absolute anchor (the value at which the signal scores exactly 50):

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

The curve flattens toward 100 as v grows. A signal with no reading contributes nothing; pillars with no contributing reading are Unrated, never zero.

Anchor table

Each anchor is the raw value at which the signal scores 50. Numbers chosen so the median agent in each population lands near 50 on each axis.

Pillar = mean of available scaled signals

For each pillar, take the scaled value of every signal in the pillar's source list that has a reading on file. The pillar score is the arithmetic mean. If no signal has a reading, the pillar is Unrated.

Headline AgentScore

weights = { adoption: 0.35, quality: 0.30, momentum: 0.20, community: 0.15 }
non_null = pillars where pillar ≠ Unrated
total_w  = Σ weights[p] for p in non_null
AgentScore = Σ (weights[p] / total_w) × pillar[p]   for p in non_null

A pillar that's Unrated drops out of the blend; the remaining weights re-normalize so the headline stays on the 0–100 scale. If all four pillars are Unrated, the agent is Unrated overall — its page shows the metadata sidebar but no composite.

Source list per kind

The formula is identical for both kinds. The source list differs because the signals that matter for an LLM and the signals that matter for an application agent are not the same — and they shouldn't be forced into the same column.

Momentum specifics

For each source where a current and a 7-day-old reading both exist:

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

0% growth maps to 50, +100% to 100, −50% to 0. If a signal first arrived inside the 7-day window, scaled_roc = 60 — a small positive bias for "newly visible".

Worked example: Claude Opus 4.7

Suppose latest signals: HN 47 mentions (7d), Bluesky 18, Reddit 12, benchmark Open LLM Average 87.6, no HF mirror.

Adoption sources (FM):
  hn_mentions_7d        scaled(47, 10)      = 80.7
  bluesky_mentions_7d   scaled(18, 10)      = 63.0
  reddit_mentions_7d    scaled(12, 10)      = 53.6
  hf_downloads_30d      no reading          → skipped
  github_stars          no reading          → skipped
  github_mentions_7d    no reading          → skipped
  → Adoption = mean(80.7, 63.0, 53.6) = 65.8

Quality sources:
  benchmark_score = 87.6  → 87.6 (no transform)
  → Quality = 87.6

Momentum (7d ROC, hypothetical +20% mention growth):
  → ~58.8

Community sources (FM):
  bluesky_mentions_7d   scaled(18, 10)      = 63.0
  reddit_points_7d      scaled(240, 100)    = 59.4
  → Community = 61.2

Headline (all four pillars present):
  0.35 × 65.8 + 0.30 × 87.6 + 0.20 × 58.8 + 0.15 × 61.2
  = 23.0 + 26.3 + 11.8 + 9.2
  = 70.3

Why some agents are Unrated

A score requires at least one signal reading on file. A model launched yesterday with no HN mentions, no HF mirror, no benchmark match and no Bluesky chatter has nothing to score. Its page renders metadata (context length, pricing, modality) but the composite is honestly absent rather than padded. As signals arrive, the pillars light up one by one.

Manipulation resistance

Three patterns trigger automatic flags: a star spike of 10× in 24 hours with low contributor diversity, a Hugging Face download surge unaccompanied by GitHub activity, and coordinated Hacker News posting bursts. Flagged signals are excluded from that day's score and the agent's record carries the reason. Themanipulation_resistanceconfidence on every score envelope reflects how clean the inputs were.

Indexes

Five indexes at launch — TAPE-100, CODE-25, WEB-25, OSS-50, MCP-25 — each with eligibility rules published in code. Equal-weight v1. Rebalances run Mondays at 03:00 UTC; every diff is logged with a short narrative explaining the largest changes.

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 therepository.