Ethos Explore

<!-- SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. -->

<!-- SPDX-License-Identifier: Apache-2.0 -->

Explore an agent's intended behavior

Capture what the agent should do. Product mission and user goals matter more than implementation inventory. The output of this skill is the data that ethos-write writes into ethos.md — the durable contract that people and agents read before they build, evaluate, review, or change this agent. Underspecified input here leads every one of them to guess at intent from code and traces.

This skill is explore first, then a mandatory intent interview. Scan the codebase and docs, infer implementation-shaped fields, then ask the user for intent source cannot own. Do not skip the interview because the scan looked complete, and do not skip a question because you think you already know the answer. Do not dump the draft ethos.md until that interview has asked at least three questions and received a reply to each.

The division of labor is: infer implementation, ask for intent. Almost everything about what the agent is can be read from source. Nothing about what the developer wants can be. Constraints, trade-offs, principles, mission, and vision live in the user's head, so they are the questions worth spending attention on.

Creating an ethos.md captures intent and authorizes no changes to the agent's application. Do not ask for optimization or change permissions, whether as a separate topic or inside another section. Permission to change code, prompts, models, tools, or evals belongs to the workflow performing that work if the user requests it. Creating an ethos.md implies no such follow-up. Preserve already confirmed organizational safety policies and their known owners, including required security review before shipping a tool with write access. Recording an existing restriction does not grant permission or require a new permission interview.

The schema you are filling

The ethos.md has five base front-matter fields, plus optional updated_timestamp, and fourteen body sections. Every body section is required. One field is a quality gate for handoff: ethos-write is blocked until Role is concrete. For any other canonical section with nothing to say, write _(none)_ rather than dropping the heading.

The fourteen headings are a floor, not a ceiling. Extra ## headings and extra YAML front-matter keys are allowed. Keep them if the user adds them. Do not strip custom sections to make the file look "strict."

Front matter

Body sections (in canonical order)

The ethos.md records durable intent, not a running list of observed failures. Track known issues and recurring failure patterns where the team tracks bugs.

Pre-flight

Check whether an ethos.md already exists for this agent. Look for ethos.md, including an ETHOS.md written by Eval Author 0.1.0, at the repository root and next to the agent's package. If one is present, ask the user whether they want to edit the existing ethos.md or start over. If they want to edit, route to ethos-write directly.

find . -maxdepth 4 -iname ethos.md -not -path '*/.git/*' -not -path '*/node_modules/*' 2>/dev/null | grep . || echo "no ethos.md yet"

If the repository has an earlier agent spec or design doc, read it as prior answers. Scan the codebase, then still run the intent interview for anything it never answered.

Step 1 — Explore the codebase

Time-box this to ~5 minutes of tool use. Read first, ask second. Greenfield projects will turn up nothing here, which is fine — move to the intent interview and ask for Role first.

1. Find agent definitions and entry points. Look for agent configs, workflow YAMLs, Python agent builders, system prompts, skills, and tool definitions:

   Copy

   find . -maxdepth 5 -type f \( -name "agent.yaml" -o -name "agent.yml" \) 2>/dev/null
   find . -maxdepth 4 -type f \( -name "*.workflow.yaml" -o -name "*.workflow.yml" \) 2>/dev/null
   find . -maxdepth 4 -type d -name "agents" 2>/dev/null

   Then use Glob / Grep to find langgraph, StateGraph, create_react_agent, system_prompt, skills, MCP servers, and tool definitions. Treat any harness or framework name you find as a clue about how the agent runs, not as a value the ethos.md must pick.

2. Find design context. Look for README.md, AGENTS.md, product/design/planning docs, launch notes, and anything in docs/. Read docs that describe goals and user value before implementation details when they look agent-relevant.

3. Map findings to schema fields. As you scan, hold a running mental table of what you can fill from the code/docs. Be honest about confidence: "inferred from system prompt" is different from "confirmed by the user."

4. Note the model only as a constraint. The ethos.md has no Model section. The agent's config already carries the model in use, and it changes without touching the ethos.md. If the scan shows a provider allowlist or pinned model family, carry it into Constraints as a draft to confirm.

   Typical inferences per field:

   • name — directory name, workflow name, or top-level package name.

   • Role — first paragraph of README, system prompt preamble, or top-level docstring. Often partial; usually needs user confirmation.

   • Purpose & Outcomes — mission from product docs, README motivation, system prompt preamble, or workflow context. Treat that as a draft to confirm, never as a skip. The outcome (the business objective) is rarely in code; even when a README names a metric, ask the user to confirm it.

   • Scope — audience from docs or prompts; categories from enumerated capabilities or named tool clusters; in/out boundaries from prompt rules.

   • Tools — from @tool decorators, tool registries, create_react_agent(tools=[...]), retrieval/corpus config, or API clients. Group low-level helpers when they share credentials, side effects, freshness, and failure modes.

   • Harness — infer from how the agent actually runs: adapter configuration, workflow YAML, service entrypoints, CLI commands, Dockerfiles, notebooks, or deployment configs. Capture behaviorally relevant capabilities, not a catalog name. If you cannot see how it runs, write _(none)_.

   • Behavior — system prompt rules ("never give medical advice"), refusal/escalation policy, tone, accepted limitations, and non-goals.

   • Success Criteria — desired production outcomes, product goals, quality standards, escalation quality, accuracy expectations, and examples of successful behavior.

   • Trade-offs — not in the code; ask in the intent interview.

   • Constraints — mostly not in the code, but scan for partial evidence worth confirming: a pinned gateway base URL or provider allowlist, region settings, redaction or PII middleware, timeout and token ceilings, and compliance notes in docs. Treat findings as a starting draft to confirm, never as the complete list.

   • Evaluation Setup — Makefile targets, scripts, CI config, eval YAMLs, metric definitions, thresholds, and coverage notes.

   • Metric Semantics — from scorer definitions, metric names in eval configs, and telemetry field names. Fill only the entries whose meaning is genuinely ambiguous from the name. Write _(none)_ when every name is obvious.

   • Principles — not in the code; ask in the intent interview even if a prompt hints at one. A prompt is the implementation, not the intent behind it.

   • Vision — not in the code. Roadmap docs, design notes, or a README's future-work section are speculative until the user confirms them. Ask. Write _(none)_ rather than guessing.

   • Open Questions — TODOs / FIXMEs in agent-adjacent code that affect safe use or evaluation.

Step 2 — Ask intent questions

The scan told you what the agent is. It cannot tell you what the developer wants. Run a real interview for that intent before you show a draft ethos.md.

This step is a hard gate. Do not present the full draft, and do not hand off to ethos-write, until you have asked at least three questions and received a reply to each. "The scan filled everything" and "asking feels unnecessary" are not reasons to skip it.

What to ask

After the scan, split fields into two piles:

• Inferred — implementation-shaped fields you can draft from source: Tools, Harness, Evaluation Setup, Behavior copied from prompts, Metric Semantics from scorer names.

• Intent — fields source cannot own: Purpose & Outcomes, Principles, and Vision always; then Constraints, Trade-offs, and a ranked Success Criteria when the repo only has eval wiring.

Walk the intent pile. Skip a topic only when this conversation already answered it. A README, prompt, or roadmap is not a substitute for a user reply on the always-ask topics below. Keep every other unanswered intent topic on the list. Greenfield work adds Role, Scope, and whether the agent is prompt-only before that intent list.

Question count is at least three, then however many remaining intent topics still need a reply. Prefer fewer sharp questions over a long checklist, but never go below three.

Always ask these three

Code and docs can look like they already answered them. They did not, until the user says so. Ask even when you have a high-confidence draft. Confirming an inference counts; silently filling the section does not.

1. Purpose & Outcomes (business objective). Confirm the mission and the result the agent is accountable for. If you inferred a metric from a README, show it and ask whether that is the target.

2. Principles. The judgment call when Behavior runs out. Reject generic virtues. Ask what this agent should do that a careless version of it would not.

3. Vision. Where the agent is headed, or an explicit _(none)_. Do not copy a future-work section into Vision without asking. Speculative vision is the usual failure.

If Role is missing or vague, ask that first so the rest of the interview is grounded. It does not replace one of the three always-ask topics.

Ask about remaining topics in this order, skipping any this conversation already answered:

1. Constraints — hard bounds on the agent's operation, including runtime action approvals where needed.

2. Trade-offs — behavioral hard gates, then a ranking of desired outcomes.

3. Ranked Success Criteria when the repo only has eval wiring.

Do not ask which parts may be optimized or changed, including in a follow-up, checkpoint, or draft review. Keep runtime action safety requirements: for example, an agent may need approval before issuing a refund. That describes its intended behavior; it does not grant permission to change its code.

Do not collect run limits. A spend cap or experiment count for one evaluation run belongs to the tool that runs it. If the user volunteers one, keep the standing policy and drop the number.

How to ask

Follow this Q&A pattern:

• One question per message. If a topic needs more depth, ask a follow-up in the next message. Do not batch four intent questions.

• Prefer multiple choice. Ground the options in what you found. Always include a way to reject the list (Something else or I don't know). Open-ended is fine when a lettered list would fake certainty.

• Wait for the reply before the next question.

• Accept "I don't know" and move on. Write _(none)_ and record the gap in Open Questions. A fabricated constraint is worse than a missing one.

• Confirm inferred bounds. If the scan found a pinned gateway or redaction middleware, show it and ask whether it is a real boundary.

• Confirm inferred intent. If the scan produced a plausible Purpose & Outcomes, Principles, or Vision, show the draft as an option. Do not treat that draft as the answer.

Example (one message, then stop):

The README says this agent exists to "cut ticket volume." Is that the business objective I should record in Purpose & Outcomes?

A. Yes — ticket volume is the outcome (tell me the target if you have one) B. Close, but the real objective is … (tell me) C. No business metric — internal tooling D. I don't know

A later Constraints example:

The scan shows a pinned model gateway and no cost ceiling in docs.

Which hard bounds should Constraints record?

A. The agent must use that gateway; no other bounds B. That gateway is required, plus a production cost or latency ceiling (tell me the number) C. No hard bounds — write _(none)_ D. Something else

Reject generic virtues for Principles. "Helpful, harmless, and honest" gives a downstream reader nothing. Ask for the judgment call: what does this agent do that a careless version of it would not? If nothing comes back, write _(none)_.

Push once on a non-answer for Trade-offs. "Balance quality and cost" is not decidable. One concrete follow-up usually produces a ranking. If it does not, record the ambiguity and move on.

If the user provides outside context, read it and update the inferred draft before the next question.

Red flags — stop and ask

These mean you skipped the interview:

• You are about to paste the full ethos.md and have asked fewer than three questions

• You filled Purpose & Outcomes, Principles, or Vision from the scan without a user reply

• Several intent questions in one message

• Handing off to ethos-write with fewer than three Q&A replies in this conversation

Step 3 — Present the draft ethos.md

After the interview, present the entire ethos.md at once — every field, with inferred values shown inline and remaining gaps as _(none)_. Then ask one question:

"Here's the full ethos.md I'd write. Tell me what to change, and I need a concrete Role before I can hand off to ethos-write."

Show the rendered ethos.md inline in markdown (one ## section per field, same shape as the on-disk file). For fields you defaulted, note the default in parentheses so the user knows they can override:

• Tools: Prompt-only. (default — say so if the agent needs tools)

• Purpose & Outcomes / Success Criteria inferred from implementation (say so if there is outside context to incorporate)

When the user could not answer a section, write _(none)_ and add the gap to Open Questions. Say "Constraints: (none) — operational limits remain unspecified" so the consequence is visible while it is still cheap to fix. An unspecified limit grants no application-change permission.

Do not walk the schema field by field in this review. Do not restart the intent interview here unless the reply surfaces a contradiction.

Do not use public-facing shorthand like AUT or "agent under test" in the rendered ethos.md. Use "this agent" for the agent being specified. Use "target agent" only where the agent's purpose is explicitly to inspect or modify another agent.

Allowed follow-ups after the draft:

1. The hard-required Role is missing or vague — ask for it before handoff.

2. The user's reply surfaces a contradiction that needs one targeted clarification (for example they say "drop the search tool" but the codebase shows the agent depends on it).

Step 4 — Hand off

After the user's reply, apply the corrections and check the one hard precondition:

1. Role is a concrete one-sentence answer (not "help with stuff").

If it is still unresolved, ask for it in one final message and stop until the user provides it. Do not hand off with Role blank — the artifact is useless downstream even though the parser accepts it.

Every body section heading must be present before handoff. Write _(none)_ for honest gaps, and note them in Open Questions when they affect safe use or evaluation. Do not drop a heading.

If the Role quality gate is satisfied, announce the handoff in one line ("Handing off to ethos-write to write ethos.md") and trigger it.

If the user pushes back

• They want to change one or two fields. Apply the edits, re-show the changed sections only, ask "good now?", proceed.

• They want to redo the whole thing. That usually means the codebase scan got something fundamentally wrong. Re-scan with their correction in mind, then re-present once.

• They keep changing their mind on Role. Stop. Tell them the agent will not be useful until they can write one concrete sentence and offer to come back later. Do not loop on rewording.

• They skip remaining intent questions. You already asked at least three, including Purpose & Outcomes, Principles, and Vision. Treat skipped topics as I don't know: write _(none)_, record them in Open Questions, and say once what it costs. Then present the draft. Do not re-open the full interview.

• They say "just write it" before three questions. Ask the next always-ask question anyway. After they answer or decline each of the three, continue.

Gotchas

• "You decide" means commit to the default and announce it. Example: "I'll record Tools: Prompt-only. Tell me to change it if not." Never silently fill in.

• Tool over-spec is the most common error. Users ask for a search tool when prompt-only would work. Probe: "Do you have evidence the model alone fails on these?" If no, drop the tool.

• Tool and harness inventory should be compressed. Do not create one row per helper method when several helpers share the same source, credential, side effect, freshness, and failure mode. Group them and call out only the differences an evaluator or reviewer needs to know.

• Mission before mechanics. An ethos.md that only says how the current code is wired is not good enough. If goal context cannot be found in the codebase or docs, say the mission is inferred from implementation and still ask the user to confirm Purpose & Outcomes. A guessed business objective is worse than _(none)_.

• Never infer a constraint you cannot verify. A guessed provider allowlist or cost ceiling misstates the agent's operating limits. Ask, or write _(none)_ and record the gap in Open Questions.

• Trade-offs make competing outcomes decidable. A priority order tells the agent and its evaluators which outcome matters when quality, latency, cost, or reliability conflict. It grants no optimization permission.

• "No behavior constraints" usually means "I haven't thought about it." Probe once: "Anything that should never appear — names, phone numbers, competitor mentions?" One probe, then move on.

• Do not skip the codebase scan even when the user seems eager to answer questions. Spending the first five minutes reading makes the interview shorter and sharper. Asking something the codebase already answers loses trust immediately.

• Do not skip the intent interview even when the scan looks complete. Ask at least three questions. Always ask Purpose & Outcomes, Principles, and Vision, even when you think you already know. One at a time. Multiple choice when you can.

• Do not invent Vision from a backlog. A future-work bullet is speculative until the user confirms it. Ask, or write _(none)_.

• A framework import is not a harness. Describe how the agent runs. Do not map an import onto a harness name. Imports such as langchain, langgraph, crewai, autogen, or pydantic_ai say nothing about that.

• Record intent, not runtime config. Describe the desired harness behavior and artifacts without asking the user to design a raw runtime config. Config files stay with the agent's implementation.

• Do not invent Known Issues fields. Known issues and recurring failure patterns belong in the team's issue tracker, not in the ethos.md.