The frontier agent harness, with context and capabilities no outside agent has.
Wisp is your personal assistant in Geist, with a team of specialist agents behind it. Tool search, skills that load on demand, threads that never run out, long goal runs, a code sandbox and review for every change: the patterns you know from Claude Code and Codex, working on your process models, maps and data.
Wisp works inside Geist, not on its screen
Your whole company as context
Process models, sites, suppliers, org data and the history of every change, read directly instead of scraped from a page.
Typed tools, no guessed clicks
Wisp calls the same tools the app uses, so it never has to guess what a button does.
Your rights, checked on every call
It works with the permissions of the person it acts for, and nothing beyond them.
Your review before anything changes
You choose whether it only reads, proposes changes for review or writes directly.
The most dangerous fact is the one nobody dated
A cargo ship searched for chemicals that were never on board. An embassy struck because the map in use was years out of date. When a system cannot tell how old its information is, it acts on yesterday with today's confidence. Wisp knows the age of every fact it uses and says so before you decide.
- Every fact keeps its observed, published and verified dates; reading it today never makes it newer
- Evidence is rated current, stale, unknown or conflicted, and a missing date stays unknown
- Before a change is applied, it is checked again against the current state
Your personal assistant, with a team of specialists behind it
Wisp can use everything on this page: it models, analyzes, researches, builds documents and runs long goals. Bigger requests become a visible plan, a lead agent hands each step to a specialist and the steps run in parallel. A failed step is replanned and the run continues. Every role keeps its own thread, compacted as it grows, so no conversation has to start over.
- Pick up a specialist's thread weeks later, where you left it
- @mention a specialist to ask it directly
- Wisp can stop and ask you a question
It remembers what was done, and why
Before Wisp touches a process or project, it recalls what changed there, who changed it and why, its own earlier work included. Memory and preferences live at three levels: your organization, your team or department, and you. Long threads fold older turns into a checkpoint and keep going.
- Organization rules apply to everyone, team preferences to one department, yours only to you
- Memory you opt into is consolidated nightly or weekly: duplicates merge, outdated entries are replaced
- It compares versions and explains a change; run /compact yourself at any time
370+ typed tools and 55 skills, loaded when a task needs them
The tools cover modeling, analysis, Horizon, event logs, org data, web search and your knowledge spaces, in 19 domains. The agent searches the catalog and loads only what the task needs. Skills work the same way, and each turn lists the skills it read.
- Your admins write skills for your own processes
- Each agent sees only the tools its role allows
- Answers from the web and your documents cite their sources
/goal keeps working until the evidence is in
A goal run holds one objective across many rounds. You cap rounds, tokens, cost, time, tool calls and changes. The run ends when a budget is spent or when it can show evidence that the goal is met.
- Press Enter while a turn runs to queue your next message, and edit the queue before it sends
- On supported models, Ctrl+Enter or Cmd+Enter steers the running turn from its next model step
- 40+ slash commands in the composer
Documents that look designed and check their own numbers
A typical chat export is a white slide with three vague bullets and figures written from memory. Wisp calculates figures in Python from your data, in a sealed container with no network, a read-only filesystem, no root user and memory and time limits. It lays out PowerPoint, Excel, Word and PDF files in a consistent design, renders every slide and checks it before you get the file.
- Each slide names the data and period its figures come from
- Org admins enable code execution, and each user decides whether to use it
- Generated files arrive in the chat as document cards
Suggestions arrive where you work
Answers arrive as interactive cards: change reviews, decisions, questions, charts, timelines, comparisons, RACI matrices and documents. Agents attach suggested changes to chat messages and notifications. You apply all of them, pick a subset, deny with a reason or discuss.
- Applying checks each change against the current model first
- Suggested elements are flagged on the BPMN canvas, and the Horizon map has an Agent Suggestions lens
- Notifications can carry an image of the affected location or flow
You decide what it may change
Three permission modes. In Read, Wisp answers and suggests, and your model stays as it is. In Review, every change batch lands on the canvas as a highlighted diff with a review card: keep it, undo all of it or undo single steps. In Write, changes apply directly and each one is recorded in the history. Destructive actions always need an explicit confirmation, and if that check cannot run, the action does not run either.
- Review cards offer a summary, a side-by-side view and the XML
- Version history with an AI summary per version, compare and restore
- A spend ceiling per thread, a Stop per run and an emergency pause for the organization
- Added
- Unchanged
Add a 48-hour timer and a reminder task, so requests stop waiting on approvers
- +Add timer “48 h without approval”
- +Add task “Send reminder”
- +Connect to “Approve request”Undo stepUndone
One thread, three levels of detail
Focus shows the answer with its result cards. Detail adds every step Wisp took to get there. Audit shows the tool calls behind each card, failed calls and retries included, and the receipt for the turn. Most people stay in Focus or Detail; auditors switch to Audit.
- Every run keeps a turn receipt and an execution trace
Why did approvals slow down in September?
38% of September requests waited on one approver for more than 3 days.
38% of September requests waited on one approver for more than 3 days.
- Read the event log for August and September
- Compared versions 12 and 14
- Found the approver bottleneck
- Drafted 3 changes
- read_event_logPurchase approval · 1 Aug–30 Sepok
- compare_versionsv12 → v14ok
- validate_model3 changesfailed
- validate_model3 changes · fixed gatewayok · retried
Strong models where it matters, cheaper ones where it doesn't
Admins decide which model runs which work. Map the pro, standard and fast tiers to any of 10 providers and 20+ models, set reasoning effort, and give each agent role its own tier: deep analysis on a frontier model, routine checks and summaries on a fast one. Switching models mid-thread keeps the history.
- Your own keys, stored encrypted per organization
- A spend ceiling per thread
- Model preferences per agent role
Switching keeps the thread · 14 messages
Your favorite agent, guided through Geist
Keep working with the agent you already use. Through WebMCP, now in beta, a browser agent finds Geist's typed tools in the open tab and works with your permissions and the same review as Wisp. When a task needs Geist's own context, it can hand the task to Wisp. Browser agents stay in Review mode unless your organization allows more.
- Next: a hosted MCP endpoint with OAuth sign-in for Claude Code, Codex and other MCP clients
- Planned: skills over MCP, so outside agents learn how Geist works from our own guides
- Planned: connections to your own MCP servers
What the harness does today
Tools and skills
- 370+ typed tools in 19 domains, loaded on demand
- 55 built-in skills, plus skills your admins write
- Answers from the web and your documents with cited sources
- Tool access limited per agent role
Context and long runs
- Recall of what changed in a process, and why
- Memory and preferences per organization, team and person
- Threads per role that never run out
- /goal runs with budgets and evidence
- Queue with Enter, steer with Ctrl+Enter
Review and audit
- Read, Review and Write modes
- Keep, undo all or undo single steps
- Version history with AI summaries
- Focus, Detail and Audit chat views
- Confirmation before destructive actions
Evidence and reliability
- Observed, published and verified dates on every piece of evidence
- Current, stale, unknown or conflicted ratings
- A failed tool call or handoff does not end the run
- Answers continue past the output limit
- Model tiers per agent role, chosen by admins
- Outside agents through WebMCP (beta)
Compare and go deeper
- Enterprise AI agents compared: harness depth in 2026A source-cited look at how in-app agents in enterprise software are built.
- Wisp and the AI TeamScheduled checks, research runs and the review flow for agent work.
- Delegation and permission checksAgents act with the rights of the person behind them, checked on every operation.
Put the harness on one of your processes
Start with a goal run or a Review session on a process your team runs today.