Talking to AEGIS

The Conversations page is AEGIS’s “ask anything” console. Type a question about your portfolio — operations, spacing, flaring, compliance, data — and AEGIS recruits the right specialist skill for the topic automatically. There is no agent to pick: typing is the only decision.

For filing preparation, the recommended workflow starts from the Compliance Dashboard rather than Conversations. The workspace-based approach provides a structured, step-by-step experience. Use Conversations for ad-hoc questions and exploratory analysis.

The Canvas layout

Conversations opens as a single centered thread — the Canvas. Everything else slides in when you need it:

Header — history (☰), the conversation title (click ✎ to rename), delete, the Context toggle, and + New.
History drawer (left) — your conversations as one recency-sorted list with search and filter chips: All, Needs review (conversations paused for human approval — the ☰ button shows an amber badge when any are waiting), and Archive (conversations from the retired specialist picks, kept readable with their original names and colors).
Context panel (right) — transparency about what AEGIS is using and spending. Covered in detail below.

Starting fresh: with no conversation selected you get a greeting, a large composer, and suggestion chips — when your portfolio has overdue items, upcoming deadlines, or flagged anomalies, the chips show those live counts. Just start typing — the conversation is created when you send, and its title is set from your first message.

Mentioning entities: type @ followed by a few letters (e.g. @mitch) and pick a well, lease, or facility from the live suggestions — the name drops into your message. You can also find any conversation by title from anywhere with ⌘K.

The composer

A multiline input: Enter sends, Shift+Enter adds a newline. Below the text sits a row of control chips:

Control	What it does
Thinking	Auto (the model decides when to reason harder) or Off (fastest). Your tenant’s Platform Settings set a ceiling; a request can lower the mode but never raise it past the ceiling.
Sources	Mirrors the data-source scope shown in the context panel (read-only in this release).
Attach	Attach files to your message — see below.
Mic	Dictate your message with speech recognition.

Attaching files

Click Attach (or drag a file onto the composer, or paste one) to include documents with your message. AEGIS reads the text out of each file when you attach it, so problems — an unreadable PDF, an empty document — show up immediately on the file pill, not after you’ve written your question.

Supported: PDF, CSV, XLSX, DOCX, TXT — up to 10MB per file, up to 3 files per message. Images aren’t supported yet.
Very large documents are truncated at a reading budget; the pill shows a truncated marker when that happens.
The agent sees the extracted text alongside your message and can answer about it in the same turn; the file stays available as context for follow-up questions in the conversation.
Your message bubble shows a pill for each attached file, including after a reload. Remove a pending file with the × on its pill before sending.

Understanding responses

While AEGIS works you’ll see status text (“Loading memory context…”, “Executing tools…”) and blue tool-activity lines (“Querying knowledge graph for offset wells…”). Answers render as full-width prose; your messages stay as compact bubbles on the right.

Under each completed answer is a metadata row:

Run chip — e.g. 2 skills · 18.4k tok. Click it to open the context panel scoped to exactly that answer’s run.
Thinking badge — appears when extended reasoning was off for that run.
Thumbs up / down — rate the response (see below).

Charts

Ask for a chart of any data the agent has assembled — “show me a chart of daily flare incidents”, “plot flared volume by month” — and the answer includes a rendered chart card (bar, line, area, doughnut, or scatter). Charts follow your light/dark theme, persist in the conversation, and rerender on reload. Any agent can chart; no skill needs to be loaded.

HITL review

If the agent prepares a filing that requires human review, the conversation pauses with an amber “Filing package ready for review” banner, and the context panel opens automatically showing the paused run. Click Review on the banner to go to the filing queue. Filing assembly is always review-gated, no matter which skill produced it.

The context panel

Open it three ways: the Context button in the header, clicking a run chip under an answer, or automatically when a run pauses for review.

It has three sections:

Data sources — what AEGIS reads when answering: the knowledge graph and production data today, with RRC filings archive and SCADA shown as not yet connected. (Source selection is informational in this release.)
Referenced entities — the wells, leases, and facilities the run actually looked up, as chips. Click a chip to open that entity’s detail page.
This run — the thinking mode, which skills were injected, and the run’s token count and cost.

Run details are captured live while you watch a turn complete. If you reload the page or come back to an old conversation, the panel shows ”—” for past runs — the answers themselves are always preserved.

Pinning: on wide screens (≥1440px) a pin button docks the panel into the layout so it stays open beside the thread instead of floating over it. The pin is remembered on this browser.

Rating agent responses

Every completed answer shows thumbs-up / thumbs-down icons. Your feedback becomes a test case that helps improve the agents.

Thumbs up — the response was good. One click; you’re done.
Thumbs down — the rating saves immediately, and an optional panel opens where you can add what the agent should have said, a note, and a category tag. Details help, but they’re never required.

You can change your rating later (icons stay when you reopen a conversation), and each person’s rating is tracked separately. Icons appear only after a response has fully finished — you can’t rate a response mid-stream.

Projects

A project is a named container for related work: its conversations, uploaded reference files (“project knowledge”), optional instructions, and default settings. Open one from the Projects section of the history drawer, or follow a Conversations / Projects / … breadcrumb.

Knowledge — up to 5 files per project (same types and 10MB limit as message attachments). Every conversation in the project can answer from them without re-uploading; the text is added once per conversation, the first time you ask.
Instructions — free text that applies to every conversation in the project (e.g. “Cite lease and API numbers in every answer”).
Defaults — currently the thinking mode: a project can turn thinking down for all its conversations (never above your tenant’s ceiling). The effective mode shows in the context panel’s “This run”.
New conversations started from the project home belong to the project. Type directly into “Ask within this project…” on the project home to start one with your question already sent. Deleting a project keeps its conversations (they return to your Recent list) but removes its knowledge files.

Building skills

Skill authoring lives on the Skills admin page — the Build with Creator button there opens this same chat surface connected to the Skill Creator. The Conversations page itself always talks to the General assistant.

Tips for effective conversations

Be specific about the entity. Include the well name, API number, or lease name so AEGIS can look up the right data — the context panel’s Referenced entities section shows you what it found.
State what you need clearly. “File a Rule 37 spacing exception for Mitchell Ranch 1H” beats “I need help with a filing.”
One task per conversation. Start a new conversation for each distinct filing or analysis to keep context clean.

Example prompts

“Which leases are close to their flaring authorization limit this month?”

“File a Rule 37 spacing exception for Mitchell Ranch 1H (API 42-329-40001-0000)”

“What changed in our compliance posture this week?”

“Is Howard Unit A-1H within its flaring authorization?”