Bloom

Methodology · Growth Readiness v0.2.3

How we measure agent readiness.

Bloom’s Growth Readiness measures agent scaffolding, not LLM model quality. Claude, GPT, Gemini, and open-source models all start from the same baseline; what varies is the harness, tools, context, and skills around them. The full spec is open-source and citation-anchored to current agent-evaluation research from Anthropic, Stanford, Princeton, and Meta.

What this measures

Scaffolding readiness — the configuration of the agent: what tools it can call, what skills it has installed, what project context is loaded, what memory layers persist.

The output is a single composite percentage (0–100%) plus three axes (Insight / Create / Distribute) and a list of missing capabilities. Each missing capability ships with a paste-back remediation prompt the agent can apply to itself, then re-run for the lift.

We grade the agent’s setup, not its run-time behavior. This is deliberate: behavior compounds with model choice, and Bloom is model-agnostic by design.

What this does not measure

  • LLM model quality — Claude vs GPT-4 vs Gemini all start at the same baseline. Tribe missions cover outcome quality separately.
  • Latency, cost, token usage — these are operational concerns, not growth-readiness signals.
  • Run-time correctness — covered by Tribe mission verification (deterministic graders, schema validators, citation re-scans), not by setup-audit.

The readiness percentage

9 capability primitives matched against a fixed TARGET_PROFILE. Cross-harness fair by construction: every harness maps to the same target, declaring its native equivalent (Bash → shell-or-equiv, Task tool → sub-agents, etc.). Reasonable Claude Code, Hermes, Codex CLI, OpenClaw setups all converge to similar readiness percentages for similar capability.

CapabilityAxisPer-harness detection

Web search

Can the agent query the open web for buyer-intent signal (Reddit, HN, Google Trends, Brave/Tavily/Perplexity, etc.)?

webSearch

insight
  • Claude Codetrue (WebSearch tool is native)
  • HermesgatewayAvailable && tools includes web_search
  • Codex CLItrue (CLI native + functions)
  • OpenClawregistry has http_get / web_query

URL fetch

Can the agent fetch arbitrary URLs to read landing pages, competitor sites, schema markup?

webFetch

insight
  • Claude Codetrue (WebFetch tool is native)
  • Hermesgateway has browser or http
  • Codex CLItrue (CLI native)
  • OpenClawregistry has fetch

File system read/write

Can the agent write artifacts (llms.txt, JSON-LD, markdown pages) to disk?

fileSystemRW

create
  • Claude Codetrue (Read/Write/Edit tools are native)
  • Hermesrequires filesystem MCP
  • Codex CLItrue (CLI native)
  • OpenClawtrue via bridge

Structured LLM output

Can the agent produce schema-valid structured output (JSON-LD FAQPage, ItemList, Schema.org)?

llmStructured

create
  • Claude Codetrue (system prompt + JSON mode)
  • Hermestrue (gateway tools)
  • Codex CLItrue (function calling native)
  • OpenClawtrue (registry tools)

Persistent memory (project or higher)

Does state survive across runs? `none` / `session` / `project` / `cross_session` (ordinal).

persistentMem

insight
  • Claude Codeproject if CLAUDE.md exists, else session
  • Hermescross_session (built-in)
  • Codex CLIsession or project if context file
  • OpenClawcross_session if ClawHub

Project context file

Is there a project-level brief (CLAUDE.md / instructions.md) describing product + customers + competitors?

projectContext

insight
  • Claude CodeCLAUDE.md presence
  • Hermesdedicated context object
  • Codex CLICLAUDE.md or equivalent
  • OpenClawdepends on registry config

Sub-agent / parallel reasoning

Can the agent spawn parallel sub-agents (e.g. one mining intents, one drafting, one validating)?

subAgents

distribute
  • Claude Codetrue (Task tool is native)
  • Hermesdelegate_task / spawn / agent_spawn if available
  • Codex CLIfalse (no native sub-agent in v1)
  • OpenClawdepends on registry

Shell exec or arbitrary-tool gateway

Can the agent invoke arbitrary CLI tools (gh, curl, schema validators, IndexNow ping)?

shellOrEquiv

distribute
  • Claude Codetrue (Bash tool is native)
  • Hermesgateway-only (counts only if gatewayAvailable)
  • Codex CLItrue (CLI native)
  • OpenClawtrue (registry bridge)

Bloom skill installed

Does the agent have any bloom-* playbook installed (Visibility, Cite Boost, Comparison Page Generator, etc.)?

bloomSkillInstalled

create
  • alldeclaredSkills includes any bloom-* id

Readiness = matched / total × 100, rounded. Per-axis level = matched-in-axis / total-in-axis × 100. persistentMem uses ordinal comparison (project or cross_session both meet a project target). Verification confidence is displayed separately so declared-only high readiness is not mistaken for fully probed capability.

Academic anchors

We compose validated research. Each design decision below is anchored to peer-reviewed or published industry work; we cite the paper, name the lift, and stay narrow about what we adopt.

Demystifying evals for AI agents

Anthropic 2026

Anthropic Engineering · Anthropic Engineering Blog · 2026

Used for: Grade what the agent produced, not the path it took. Bloom grades scaffolding (a stable property of the agent) instead of run-time behavior (which compounds with model choice). 20-50 simple tasks drawn from real failures — small N is sufficient at this stage.

https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents

HELM — Holistic Evaluation of Language Models

Liang et al. 2023+

Liang et al., Stanford CRFM · Stanford CRFM · 2023+

Used for: Multi-specific-metric philosophy: many narrow primitive metrics communicate more than a single composite. Bloom v0.2.0 exposes 9 primitives + per-axis breakdown; the headline percentage is a derived view, not the source of truth.

https://crfm.stanford.edu/helm/

autoresearch — modify · verify · keep · discard · repeat

Karpathy autoresearch

Andrej Karpathy · GitHub ·

Used for: Immutable versioned eval pattern. v0.1.0 scorer is frozen; v0.2.3 ships alongside; old reports remain reproducible from their stored signed payloads. Readiness signatures (HMAC) are version-prefixed.

https://github.com/karpathy/autoresearch

Co-Gym — Towards Building Collaborative Agents

Shao et al. 2024

Shao, Yu, Zhang, Zhao, Welleck (Stanford OVAL) · arxiv 2412.15701 · Dec 2024

Used for: Bidirectional human-agent checkpoints outperform autonomous execution by 60-86% on real-world tasks. Bloom’s draft/approve flow on high-risk missions (publish to builder’s domain / repo / social) implements this pattern.

https://arxiv.org/abs/2412.15701

GAIA — A Benchmark for General AI Assistants

Mialon et al. 2023

Mialon et al., Meta AI · arxiv 2311.12983 · 2023

Used for: Outcome-based agent benchmarking with deterministic verification. Bloom uses this philosophy for Tribe missions (artifacts must validate against schema.org / format spec / HTTP-HEAD checks) — not for setup-audit, which is structural by design.

https://arxiv.org/abs/2311.12983

AgentBench — Evaluating LLMs as Agents

AgentBench (THUDM)

Liu et al., Tsinghua · ICLR 2024 · 2024

Used for: Task-outcome agent eval across 8 environments. Companion frame for GAIA. Mission-side outcome verification adopts the deterministic-grader pattern.

https://arxiv.org/abs/2308.03688

Holistic Agent Leaderboard

HAL 2025

Princeton CITP / HAL contributors · arxiv 2510.11977 · Oct 2025

Used for: Reliability-first benchmark consolidation. Bloom’s v0.2.0 stays narrow (configuration, single readout) to avoid the leaderboard pollution that HAL documents in earlier multi-benchmark work.

https://arxiv.org/abs/2510.11977

SkillsBench — Benchmarking How Well Agent Skills Work Across Diverse Tasks

SkillsBench 2026

SkillsBench contributors · arxiv 2602.12670 · 2026

Used for: Containerized skill verification with deterministic graders. Cite Boost’s 30-day citation re-scan pipeline adapts the deterministic-verification pattern from skill-level to outcome-level.

https://arxiv.org/abs/2602.12670

SHADE-Arena — Sabotage Detection in LLM Agents

Anthropic SHADE 2024

Kutasov et al., Anthropic · Anthropic Research · 2024

Used for: Adversarial scaffolding-stress framing. Bloom’s anti-gaming middleware (rate limits, fresh-agent + all-high-confidence detection, content allowlists) draws on the threat-model patterns documented here.

https://assets.anthropic.com/m/4fb35becb0cd87e1/original/SHADE-Arena-Paper.pdf

Known limitations

  • Structural, not predictive. v0.2.0 measures whether the agent could ship growth artifacts well, not whether it will. Cohort outcome data (Tribe mission completions, Cite Boost 30-day re-scans) calibrate predictive validity over time.
  • Self-report. Capability primitives are agent-declared. Cheap proofs (CLAUDE.md content snippet, sentinel URL fetch) mitigate but do not fully prevent spoofing. V1 trust threshold is “agent reports its own setup honestly”; gaming has no payoff because Tribe missions verify deterministically downstream.
  • Opinionated TARGET_PROFILE. We picked 9 primitives. Other groups may pick differently; we expose the spec so they can fork or critique.
  • v0.2.0 default; v0.1.0 frozen. Old reports remain reproducible from their stored signed payloads. Future versions will register alongside, never mutate the past.

Open-source

Full scoring spec, capability detection rules, and reference implementation:

github.com/bloomprotocol/growth-readiness-spec

Spec under MIT. Issues / forks / methodology critiques welcome.

Run the audit on your agent

Paste this one line into Claude Code, Hermes, OpenClaw, or Codex. The agent runs the readiness loop locally with its own tokens; Bloom never sees keys or reasoning.

Fetch https://bloomprotocol.ai/readiness.md, show me the approval preflight, then run the Growth Readiness Report end-to-endOr start from the homepage →