Static linter for AI agent configs, system prompts, and tool definitions. 7 structural detectors (H1–H7), 6 HERM v1.1 scoring dimensions, validated against 28 comparison files. 154 tests (including a CI-mechanical doc-consistency gate), 0 LLM calls per scan, ~2ms per file. Reproduce: bash evals/sample-detection-rate.sh flags 4-of-4 known-bad samples and passes 1-of-1 clean — same input, same output, every run.
AI agent configs fail for language reasons long before they fail for code reasons: vague tool descriptions, missing stop conditions, and schema fields that say nothing useful.
lintlang catches those language-level failures before they hit CI, runtime, or human review — without calling a model.
- "My agent picks the wrong tool because the tool descriptions all sound the same."
- "We only catch prompt and config drift after the agent starts looping."
- "I want a prompt linter or agent-config linter that runs in CI with no model calls."
- "Our YAML is valid, but the instructions inside it are still bad."
pip install lintlanglintlang scan samples/bad_tool_descriptions.yamlLINTLANG v0.2.1
samples/bad_tool_descriptions.yaml
FAIL — 1 CRITICAL, 2 HIGH, 6 MEDIUM, 3 LOW
H1: Tool Description Ambiguity
[CRITICAL] tool:process_ticket
Tool 'process_ticket' has no description.
Most agent-config "review" tools call an LLM to grade your YAML. That makes the review expensive, slow, and itself non-deterministic — the same config scores differently on Tuesday versus Thursday. lintlang skips the model entirely.
| LLM-based config review | lintlang | |
|---|---|---|
| Cost per scan | $0.01–$0.50 (model + tokens) | $0.00 |
| Wall time per file | 2–15 s | ~2 ms |
| Same input → same output | No (sampling-dependent) | Yes (regex + AST) |
| Runs offline / in CI without keys | No | Yes |
| Catches vague tool descriptions | Sometimes | Always (H1) |
| Detects missing termination conditions | Rarely | Always (H2) |
Detection rules are static regex + structural heuristics. The same input produces the same output, every run, every CI.
Use lintlang when you author or review AI agent tool descriptions, system prompts, or config files and want a static prompt/config quality gate in CI before runtime testing.
- Semantic correctness — lintlang is structural. It catches vague tool descriptions, not wrong ones. ("delete_user" with empty description fails; "delete_user" pointing at the wrong table is invisible to lintlang.)
- Open-ended creative writing — H1–H7 are calibrated for agent configs and system prompts, not prose.
- Auto-fix — lintlang reports findings; it doesn't rewrite. Pair with a human or LLM for the fix step.
- Behavioral safety proofs — a clean lintlang scan is a necessary but not sufficient condition for agent safety. Run a runtime evaluator (e.g., the rest of the Hermes Labs audit stack) for dynamic checks.
- Config formats we don't parse yet — currently JSON, YAML, plain text, and
.prompt. Markdown front-matter parses; arbitrary nested templates may not.
Static linter for AI agent tool descriptions, system prompts, and configs.
Most AI agent bugs aren't code bugs — they're language bugs. Vague tool descriptions make agents pick the wrong tool. Missing constraints cause infinite loops. Schema mismatches break structured output. lintlang catches these at authoring time, in CI, with zero LLM calls.
pip install lintlangRequires Python 3.10+. One dependency (pyyaml). No API keys, no network access, no LLM calls.
# Scan a single file
lintlang scan agent_config.yaml
# Scan a directory (finds .yaml, .json, .txt, .md, .prompt)
lintlang scan configs/
# JSON output for CI
lintlang scan config.yaml --format json
# Fail CI on CRITICAL/HIGH findings
lintlang scan config.yaml --fail-on fail
# Fail CI on any MEDIUM+ findings
lintlang scan config.yaml --fail-on review LINTLANG v0.2.0
bad_tool_descriptions.yaml
──────────────────────────────────────────────────
❌ FAIL — 1 CRITICAL, 2 HIGH, 6 MEDIUM, 3 LOW
H1: Tool Description Ambiguity
!! [CRITICAL] tool:process_ticket
Tool 'process_ticket' has no description.
→ Add a specific description explaining WHEN to use this tool.
! [HIGH] tool:get_user_info
Tool 'get_user_info' has a very short description (13 chars)
→ Expand to include purpose, when to use, expected input/output.
~ [MEDIUM] tool:handle_request
Tool 'handle_request' starts with vague verb 'handle'.
→ Replace with a specific action verb.
H2: Missing Constraint Scaffolding
! [HIGH] system_prompt
System prompt defines tools but has no termination conditions.
→ Add: 'Maximum 5 tool calls per task. Stop and report after 2 failures.'
──────────────────────────────────────────────────
lintlang v0.2.0 | H1-H7 structural analysis | Zero LLM calls
lintlang gives you a verdict, not a score:
| Verdict | Meaning | When |
|---|---|---|
| ✅ PASS | Ship it | Only LOW/INFO findings or none |
| Has blind spots | MEDIUM findings present | |
| ❌ FAIL | Will break in production | CRITICAL or HIGH findings |
Each finding includes the pattern (H1-H7), severity, location, and a concrete fix suggestion. No vague "improve your prompt" — specific rewrites you can apply immediately.
These aren't arbitrary rules — they're the 7 structural failure modes that cause real agent breakdowns in production. We identified them across audits of 8 major AI frameworks (LangChain, Semantic Kernel, AutoGen, smolagents, LiteLLM, Anthropic SDK, OpenAI SDK, Agno) and 12 filed PRs. Each detector maps to a specific class of bug that no other linter catches because they're language problems, not code problems.
No existing tool covers this: yamllint checks syntax, semgrep checks code patterns, ruff checks Python style. None of them can tell you that your tool description is ambiguous enough to cause wrong-tool selection, or that your system prompt lacks termination conditions and will loop forever.
| Pattern | Name | What Users Report | Severity |
|---|---|---|---|
| H1 | Tool Description Ambiguity | "Agent picks wrong tool" | CRITICAL-MEDIUM |
| H2 | Missing Constraint Scaffolding | "Agent loops infinitely" | CRITICAL-HIGH |
| H3 | Schema-Intent Mismatch | "Structured output broken" | CRITICAL-LOW |
| H4 | Context Boundary Erosion | "Agent leaks state across tasks" | HIGH-MEDIUM |
| H5 | Implicit Instruction Failure | "Model doesn't follow instructions" | MEDIUM-LOW |
| H6 | Template Format Contract Violation | "Agent broke after prompt change" | MEDIUM-INFO |
| H7 | Role Confusion | "Chat history is messed up" | CRITICAL-MEDIUM |
H5 distinguishes between safety constraints and style negatives. Security rules like "Never expose API keys" are correctly exempted. Style issues like "Don't be verbose" are flagged with positive rewrites.
Validated on 26 real-world configs (OpenHands, RAG agents, HIPAA compliance, financial advisors, content moderation, DevOps safety) — see samples/ for examples.
Zero cost, zero latency, zero data exposure. Runs in CI where LLM calls can't. Catches structural patterns (missing termination, schema mismatches, role ordering) that LLMs are blind to because they process content, not structure.
- name: Lint agent configs
run: |
pip install lintlang
lintlang scan configs/ --fail-on fail| Flag | Exits 1 when | Use case |
|---|---|---|
--fail-on fail |
Any CRITICAL/HIGH finding | Blocking deploy gate |
--fail-on review |
Any MEDIUM+ finding | Strict quality gate |
--fail-under 80 |
Quality score < threshold | Legacy score-based gate |
# Only show CRITICAL and HIGH
lintlang scan config.yaml --min-severity high
# Only check specific patterns
lintlang scan config.yaml --patterns H1 H3from lintlang import scan_file, compute_verdict
result = scan_file("config.yaml")
verdict = compute_verdict(result.structural_findings)
print(f"Verdict: {verdict}") # PASS, REVIEW, or FAIL
for finding in result.structural_findings:
print(f" [{finding.severity.value}] {finding.description}")
print(f" → {finding.suggestion}")# Scan a directory
from lintlang import scan_directory, compute_verdict
results = scan_directory("configs/")
for path, result in results.items():
verdict = compute_verdict(result.structural_findings)
print(f"{path}: {verdict}")lintlang auto-detects file format:
- YAML (
.yaml,.yml) — OpenAI function-calling format, tool definitions - JSON (
.json) — OpenAI and Anthropic tool schemas, message arrays - Plain text (
.txt,.md,.prompt) — System prompts, instruction docs
Unknown extensions are tried as JSON → YAML → plain text.
| Tool | What It Does | How lintlang Differs |
|---|---|---|
| promptfoo | Tests prompts via eval suites at runtime | lintlang is static — no LLM calls, catches issues at authoring time |
| guardrails-ai | Validates LLM outputs at runtime | lintlang catches root causes (bad instructions), not symptoms |
| NeMo Guardrails | Runtime dialogue rails | lintlang operates on config files, not live conversations |
| eslint / ruff | Lints source code | lintlang lints natural language in agent configs |
lintlang treats tool descriptions, system prompts, and agent configs as lintable artifacts — static analysis for prose, like eslint for JavaScript.
git clone https://github.com/hermes-labs-ai/lintlang.git
cd lintlang
pip install -e ".[dev]"
pytestHermes Labs is building the reliability stack for the agent era. Memory, evaluation, observability, containment — the infrastructure layer that makes autonomous AI agents production-grade. Founded 2025 by Rolando (Roli) Bosch, solo founder, AI-amplified ("cyborg engineering"). Based in the San Francisco Bay Area.
The technical thesis: language sets the capability and intelligence; the model is the ceiling, not the source. Reliability is a question of linguistic infrastructure, not model tuning. Formalized as LPCI (Linguistically Persistent Cognitive Interface) — transfer entropy ≈ 0 in embedding-space proxy, Markov property holds, the substrate is linguistic. The engineering follow-on: when language is the substrate, the engineering is interpretive — recovering meaning across the boundaries between model and user, session and session, training and runtime.
Public technical receipts. The flagship open-source release is fidelis — zero-LLM agent memory with integer-pointer fidelity. 73.0% end-to-end QA on LongMemEval-S, Wilson 95% CI [68.7%, 77.0%], at $0 per query, fully local. Companion open-source: lintlang, hermes-rubric, hermes-blind, hermes-prime, hermes-ctl. Published research at zenodo.org and the Hermes Labs paper line. The OSS surface is the proof; the commercial work is enterprise deployments.
For enterprise deployments and AI-reliability engagements: rbosch@lpci.ai · lpci.ai
On naming. Hermes Labs is named for Hermes, the Greek messenger god — patron of communication and interpretation, the herald who carries meaning between worlds. The thread to the work: hermeneutics, the theory of interpretation that takes its name from Hermes, is the philosophical anchor for an AI infrastructure company whose substrate is linguistic. Not affiliated with NousResearch's Hermes LLM line or their hermes-agent framework — different companies, different work.
Founder: Rolando (Roli) Bosch. Site: hermes-labs.ai Citation: Bosch, R. (2026). Hermes Labs: AI reliability infrastructure for autonomous agents. https://hermes-labs.ai
Quantitative sources for claims above:
- fidelis 73.0% / Wilson 95% CI [68.7%, 77.0%]: see fidelis/README.md "End-to-end QA accuracy" + experiments/zeroLLM-FLAGSHIP-evidence/, 470 questions, eval date 2026-04-24
- LPCI thesis (TE ≈ 0 embedding-space proxy): langquant repo, commit dd918cc (2026-03-28) "LPCI PROVED" + lpci_rigorous.py:507-571
- 24-failure taxonomy: hermes-rubric/calibration/failure-mode-taxonomy.md