When to Use
Use this pattern when work spans multiple steps, a wrong action is costly, and the human owns the outcome.When Not to Use
Avoid this pattern for trivial, low-stakes, easily-undone actions. Related: Match effort to stakes, PR-05UX Pattern
Human-Initiated Interaction
1. Atlas proposes an Action Plan
The plan shows the high-level steps Atlas intends to take, each with the reason behind it. Atlas revises the plan as needed, and the conversation doesn’t progress until the human approves one. Related: Action plans, PTN-362. Atlas assembles content and reasons through judgment calls
As it assembles and reasons, Atlas shows what it’s reading, checking, and assuming. It surfaces this as a one-line summary the human can expand for detail. Related:- Show the work, PR-03
- Working transcript, PTN-16
- Assemble and Reason steps from the 5 parts
3. The situation package appears when review needs it
Atlas surfaces the read-only situation package — the context and evidence its reasoning is based on — whenever the human needs to validate that reasoning against the underlying data. Related: UX Surface: Situation Package4. Atlas proposes recommendations
Each recommendation carries what Atlas recommends, why, and the evidence it used. It offers the actions the human can take. Related: UX Surfaces: Atlas in-line message, Suggested Changes pattern5. The human decides
The human acts on the recommendation, or asks questions and requests changes. Atlas revises and re-presents, and nothing runs until the human confirms. Related:- Interrupt and steer, PTN-33
- Keep the human senior, PR-04
- Decide step from the 5 parts
6. Modify in place
If the human chooses to modify, the content appears in edit mode in the right panel (for now), pre-filled with the selected changes. The human edits fields directly, then saves or cancels. Related: Reversibility, PTN-32 UX Surface: Suggested Changes pattern7. The human confirms and Atlas acts
Atlas performs the action. Every executed action and decision stays visible for the audit trail. Related: Act step from the 5 parts UX Surfaces: Atlas in-line message, What Atlas Did pattern8. Atlas proposes the next step
Atlas keeps the conversation moving by bringing the human directly to the next step, or by suggesting next-step options. Related: Suggested actions, PTN-04Scheduled or Time-Based Trigger
1. Atlas performs a skill on its own
Atlas triggers on a schedule or an event, without a human initiating it first. Scheduled triggers run at a fixed time — a morning brief at 6am daily, an AR staging scan at 10pm on Mondays. Event triggers fire when something happens — a payment fails, an estimate sits dormant for 48 hours, a job completes.2. Atlas assembles content and reasons through judgment calls
This step stays invisible in the moment because Atlas runs in the background. Its output becomes the guidance content surfaced in the next step. Related: Assemble and Reason steps from the 5 parts3. Atlas surfaces guidance in the human’s Workspace
When surfacing guidance is part of the skill, Atlas presents one or more recommendations in the human’s Workspace. Each piece of guidance carries what Atlas recommends, why, the evidence it used, what it’s unsure about, and whether it’s one action or several sequenced steps. Atlas offers specific actions — accept, modify, defer, reject, escalate, or assign — using concrete action copy like “Approve and Send” rather than a generic “Accept.” Related: UX Surfaces: Workspace, recommendations pattern4. Some recommendations open a guided flow
For recommendations that need more detail — specific line-item changes, for example — the action brings the human into Atlas to walk through a flow. The human allows, denies, or modifies each item individually or in bulk. UX Surfaces: Atlas UI shell, Suggested Changes pattern5. The human decides
The human accepts, modifies, defers, rejects, escalates, or assigns the recommendation, and can ask Atlas for more information before deciding. Related:- Interrupt and steer, PTN-33
- Keep the human senior, PR-04
- Decide step from the 5 parts
6. The human confirms and Atlas acts
Atlas performs the action — it saves the invoice, sends the note, creates the task. High-stakes actions, like customer-facing sends or escalations, sometimes include an explicit confirmation step before execution. When the action launches inside an Atlas conversation, Atlas confirms what happened in-line. The What Atlas Did section is the person’s audit of the act: what actually changed in ServiceTitan, on whose authority, and when. When the action launches a confirmation dialog instead, the recommendation card shows as a completed step once the human confirms. Related: Act step from the 5 parts UX Surfaces: Atlas in-line message, What Atlas Did, recommendations pattern7. Atlas remembers this interaction
Atlas creates a Decision Trace for every interaction, so the person can audit a past decision and so Atlas gets better over time. The domain team authors two things for every trace:- What carries forward — what goes into Memory versus what stays in the record. Not every trace becomes a lesson: the domain team writes only governed, durable lessons to Memory.
- At what scope — customer-level (this customer always short-pays), tenant-level (shop policy), or platform-level. A lesson stays at the person’s scope unless the domain team deliberately promotes it.
- Remember step from the 5 parts
- Legible Memory, PTN-38
- Relationship Continuity, PTN-42
Do / Don’t
Do
- State each step’s concrete impact — the specific records, recipients, or amounts it touches — so a wrong step is easy to spot
- Let the human edit the plan — remove, reorder, or adjust steps — before running, not just approve or reject it whole
- Approve once at the plan level, then run the full sequence without re-asking about each step
- Stop and re-present the plan if a step fails or the approach changes mid-run — never improvise silently
- Let the human interrupt and redirect mid-run without losing completed work
Don’t
- Ask for approval on a vague plan, like “update the relevant records” — the human should know exactly what they’re approving
- Show one plan, get approval, then execute a different one
- Put trivial steps behind a full plan — it teaches people to approve without reading
- Make reviewing the plan cost more than doing the task itself
- Expose raw tool logs — summarize, and let the human expand for detail