AI Agent Map

AI Agent Map is a practical, visual-first guide for comparing mainstream AI agents, agent platforms, runtimes, and orchestration tools.

The goal is simple: help readers get to a sensible shortlist faster.

What This Repo Is For

• The agent landscape is crowded.
• Many resources explain ideas, but not fit, anti-fit, or operating cost.
• People usually need a comparison layer, not another pile of links.

This repo stays focused on selection: what a system is good at, where it breaks down, and what kind of operator cost comes with it.

Where To Start

Recent Heat Ranking

Popularity is not fit.

This table tracks projects that showed up as especially hot in the latest weekly GitHub snapshot. The rank follows the 7-day gain. The total star counts below were checked when this repo was updated.

Last updated: 2026-10-28 · Snapshot window: 2026-10-21 → 2026-10-28 (gain since last update, 7 days against last window’s 7, so the two columns compare directly for a third straight week) · Star counts: checked at update time

Project names link to the upstream GitHub repo. When this map has a written profile, it is linked separately in the “Map status” column.

• Heat is useful for discovery, not for selection by itself.
• This window is 7 days against last window’s 7, so raw gains compare directly for a third straight week; no weekly-rate conversion is applied below.
• The board cooled sharply. Only 15 of 57 measured repos accelerated and 42 slowed (last window 31 and 24). The ten seats together gained 31,952 against 44,336 (−28%), and the bar for #10 fell 19% (1,920 → 1,556). Nine of the ten seat-holders slowed; only TradingAgents sped up.
• DeepSeek Harness takes #1 for the second time, by holding still. +7,011 is down just 1% on last window, while mattpocock/skills fell 38% to +5,483 and ended a three-window run at #1 (its 15 #1s in 28 windows still lead the board; DeepSeek Harness now has 2). It is the same pattern as its first #1 on 09-30, which also came on a gain that did not rise — and that one lasted a single window. Past 260k; one pre-release (dsh-v0.2.1-rc.3, Oct 23).
• mattpocock/skills crossed 300k but did not catch Superpowers. At 304,687 against 306,144 it is 1,457 behind, down from 3,630; last window’s note that the gap would close “inside a week” at that pace was wrong, because the pace fell. And for the first time since 09-30 it gained less than the other three general collections combined (5,483 against 3,310 + 1,906 + 1,556 = 6,772) — so the “one directory is the wave” read that held for three windows does not hold this one.
• The spike rule paid out on all three of last window’s unexplained jumps. scientific-agent-skills gave back 68% and left the board (11th, 78 short of 55k); CLI-Anything gave back 56% and Ruflo 52%, though both kept a seat (#9, #8) because the bar fell too. The one documented breakout, OpenCode, also gave back 53% (+2,841, #5) and kept its seat; v2.2.0 shipped Oct 27.
• This window’s own spike is TradingAgents: +4,677, up 250%, from 14th to #3 — its best seat and largest gain since it held #1 on 05-08. The window’s only release, v0.7.1 (Oct 27), is a bug-fix release (filing dates, settlement across market holidays, SEC statement periods), and nothing first-hand ties it to the gain. By the board’s rule it is a spike until a second window confirms it.
• The skills wave holds 4 of 10 (#2, #4, #7, #10) for a third straight window, with a different fourth seat again: anthropics/skills returned at #10 (first seat since 09-30) as K-Dense left. The other six seats are four agents (#1, #3, #5, #6) and two infra projects (#8, #9). Hermes Agent (28 of 28 windows), Superpowers and mattpocock/skills (23 each) are the long runs still going; DeepSeek Harness and OpenCode are next at 8.
• The model layer moved on a new OpenAI tier, a Kimi point release and Anthropic’s corporate news, not on defaults. (1) OpenAI’s GPT-6.1 Terra (Oct 26) lists at $1 / $5 with $0.05 cached input and a 1.05M context, repricing the whole request past 272k input tokens; it reached Codex and ChatGPT Work on Plus and above the same day as a selectable model, not a default. (2) Moonshot’s Kimi K3.1 Pro (Oct 24 on Moonshot’s own index) is on its API at $2 / $10 with $0.20 cache reads (a Highspeed variant at $4 / $20); Moonshot says the weights follow by Nov 4, and its quickstart still shows unfilled [xx.x%] benchmark placeholders, so no score is recorded here. K2-series models close to new users on Oct 31 and retire by the end of December. (3) Anthropic donated Agent Skills to the Agentic AI Foundation (Oct 22), saying more than 70 agent platforms have integrated the standard, and filed an S-1 for a Nasdaq listing as “ANT” (Oct 26). (4) Anthropic also opened a waitlisted preview of Claude Duet (Oct 21), a separate personal-agent app, profiled this refresh — see personal agents. No tracked coding agent changed its default model for a fourth straight window: Claude Code v2.1.307–v2.1.311, Codex CLI rust-v0.164.0 (still bundles GPT-6.1 Sol first), OpenCode v2.2.0, Qwen Code and Kimi Code; Gemini CLI shipped no stable release. OpenCode’s Oct 17 incident now has a vendor statement on record: its X post says a breach “related to a large scale distillation attack” affected 1.4% of OpenCode Console users, the subset with external providers linked, all notified by email. Last window this map could not read it; the dated line in the timeline is updated in place. The cost table adds GPT-6.1 Terra; every other row was re-checked against the Anthropic, OpenAI, Google and Meta pricing pages and is unchanged, and both promotions (GPT-5.6 Sol to Nov 21, Gemini 4 Argon to Nov 30) are still live. Details in market-events.
• Milestones, checked against the raw counts rather than the rounded column: mattpocock/skills 300k, Superpowers 305k, DeepSeek Harness 260k, OpenCode 220k, TradingAgents 115k, addyosmani/agent-skills 110k, CodeGraph 75k, CrewAI 60k, Omnigent 11k and Kimi Code 8k. Near misses: agentmemory is 58 short of 30k, scientific-agent-skills 78 short of 55k, Aider 228 short of 50k, Ruflo 285 short of 80k, mattpocock/skills 313 short of 305k.
• OpenClaw remains the absolute leader at 393.2k stars (+557, down 21%); it is profiled but stays out of the gain-ranked table because reliable week-over-week deltas for a project this large are noisy.

Ranking Trend

How the weekly top 10 has shifted since tracking began — each line is one project, breaks mean it fell off the board that week:

And the same windows read as seats per layer — the quantitative version of the skills-wave story the bullets tell in prose:

Full stock rankings by category — agents, agent infra, skills, and their verticals, sorted by total stars — live in rankings/.

Beyond The Rank

Popularity tells you what to look at. These six pages tell you what to pick:

• Capability matrix — every project scored side by side (●/◐/○/—) across the nine shared capability dimensions, grouped by route. The answer to “for this capability, who treats it as a core strength.”
• Cost & benchmarks — frontier-model capability vs per-token price, plus how each coding agent actually bills. Since the model layer went tiered and metered, “which tier for this task” is the selection decision.
• Memory approaches — seven different things projects mean by “has memory,” from self-editing stores to passive semantic recall, and which to pick for what you need to persist.
• Consumer personal agents — the agents you hand your day to — ChatGPT dots, Gemini Task, Claude, Muse, Grok Bot, Cue, Doubao, Qwen, AutoGLM, Yuanbao, Today — compared on facts and control rather than popularity, because closed products have no public number to rank by.
• Agent Skills — what a skill is, which agents load one, and the five largest collections and frameworks (Superpowers, mattpocock/skills, ECC, anthropics/skills, addyosmani/agent-skills) compared on kind, footprint, maintainer and license, plus what to check before installing someone else’s. Collections are compared here rather than profiled.
• Observability & evaluation — the layer under everything above: once an agent runs unattended, failure stops looking like a crash and starts looking like silent quality drift. Compares Langfuse, Opik, Phoenix, Helicone, LangSmith and others — and untangles the four different things “open source” means in that field.

Market Pulse

The three structural stories shaping selection right now — full records with dates and sources live in market-events.md:

• The .claude/skills wave holds four seats for a third straight window, and its biggest directory lost #1 (May 2026 → ongoing): curated skill collections and skills frameworks have held between two and six of the weekly heat top 10 for five months, and the count rotates rather than sitting still — 6/10, then 2/10, 3/10, 4/10, 5/10, 3/10, 4/10, 4/10, and 4 of 10 a third straight window (2026-10-28). The fourth seat changed hands again: anthropics/skills (#10, +1,556) replaced scientific-agent-skills, which fell off the board. The core remains mattpocock/skills, Superpowers and addyosmani/agent-skills. mattpocock crossed 300k (304,687) but lost #1 to DeepSeek Harness after three straight windows on top, and is still 1,457 stars behind Superpowers (306,144). It also stopped out-gaining the other three general collections combined: +5,483 against their +6,772, where last window it led by 530. The key-person risk this map has flagged since June is the least lopsided it has been in weeks. This map profiles the framework end through Superpowers and tracks collections on the skill boards.
• The model layer became a budget decision — and the ladder now has every rung filled at both vendors: Claude Fable 5.1 (Sept 1) and GPT-6 Astra (Sept 3) still both list at $10 / $50; below them sit Opus 5.5 ($4 / $20, Sept 22), the two mid tiers Sonnet 5.5 and GPT-6.1 Sol (both $2 / $10, Sept 28–29, both their vendor’s coding-agent default), and now the floor: GPT-6 Luna ($0.10 / $0.50, Sept 22) and Claude Haiku 5.5 (Oct 7, $0.10 / $0.50 for prompts up to 100K tokens, $0.50 / $2.50 above). On Oct 26 OpenAI filled the rung between the floor and the mid tier with GPT-6.1 Terra at $1 / $5 (cache $0.05), selectable in Codex but not made a default. Google joined the priced field on Oct 13 when Gemini 4 Argon opened to the paid API at $2 / $10 through Nov 30, then $4 / $20, and Meta’s Muse Wonder (Oct 12) lands on $4 / $20 with $0.20 cache reads — Opus 5.5’s numbers exactly, which makes $4 / $20 a three-vendor price point. The column this map said separated the two mid tiers has closed: Anthropic’s pricing page now lists Sonnet 5.5 cache reads at $0.10/M, not the $0.20 recorded from the Sept 28 changelog (Claude Code v2.1.296 corrected its own figure on Oct 9), so Sonnet 5.5 and GPT-6.1 Sol are identical on sticker and on cache. What still separates them is long-context shape (OpenAI reprices the whole request past 272k input tokens; Anthropic does not; Haiku 5.5 prices by prompt length). See cost & benchmarks.
• Product boundaries are collapsing upward: OpenAI merged Codex into the ChatGPT app (July 9), and at DevDay (Sept 29) moved it further off the laptop with reusable cloud environments usable from any device — on the OpenAI side, “which coding agent” keeps turning into “how you use ChatGPT.” The bundled default model changed twice in a month: GPT-6 Astra from rust-v0.153.4 (Sept 4), then GPT-6.1 Sol from rust-v0.159.1 (Sept 29), carried to the main release line by rust-v0.161.0 (Oct 7). Since 0.162.0 (Oct 8) the TUI also follows the server’s model default, so the effective default can change without a release, and on Oct 14 OpenAI retired GPT-5.5 from Codex and ChatGPT for ChatGPT sign-in. Four windows have now passed with no tracked coding agent changing a default in a release. If you run Codex on defaults, pin the model id anyway — and know that a pinned model can be retired under you; the product will not hold either still for you. See Codex.

The First Cut Of The Map

Current Mainstream Coverage

98 profiled projects, grouped by what they are. Expand a group, or browse the full route/coverage tables in agents/.

Example Reading Paths

If you are still deciding where to begin, use one of these quick routes and then branch out.

Disclaimer

Star counts and 7-day gains are point-in-time GitHub snapshots taken when the repo is updated; numbers shift quickly between weekly refreshes and small rounding differences are expected. Project descriptions, vendors, and capability summaries reflect public information at the time of writing and may change as projects evolve, get acquired, or pivot. This map is selection guidance — not endorsement, financial advice, or a production-readiness guarantee. Verify against each project’s own docs before committing to a choice.