Model & settings
Any chat model enabled for your tenant, or the tenant default. Tune reasoning effort and verbosity where the model supports them — and swap models from a dropdown.
Agents built from governed parts.
An agent is a versioned prompt, an approved model, scoped knowledge and exactly the tools IT grants it. Assemble one in minutes — or describe the outcome in plain English and review the blueprint the AI Agent Builder drafts for you.
Anyapproved chat model
Per-toolMCP grants
Pinnedprompt versions
Everyrun traced
Chat model
Claude · Amazon Bedrock ▾
Reasoning effort
Medium ▾
Admissions voice + policy · v9 pinned
Knowledge & websites
✓Admissions policies✓Course pages · max 8 · ≥ 0.55✓HR handbook
MCP tools · Gecko Engage
✓findContact✓findEvents✓addContactToSession✓sendEmailToContact
Tool choice
Auto
Max tool steps
6
Anatomy of an agent
Every part of an agent is a governed choice.
Builders assemble agents from components IT has already approved. Nothing is granted by default — not a model, not a folder, not a tool.
Prompt
A saved prompt from the central library, with variables, logic and reusable fragments — optionally pinned to a specific version.
Knowledge
Only the knowledge folders and websites you select. Per-site limits of 1–20 results and a minimum relevance score from 0 to 1 filter weak matches before the model sees them.
MCP tools
Per-server, per-tool checkboxes. An agent gets exactly the tools an admin grants — never a whole server by accident.
Provider-native tools
Where the model supports them: OpenAI web search (domain-restricted) and file search; Anthropic web search, web fetch and code execution.
Tool behaviour
Set tool choice to auto, required or none, and cap how many tool steps a single reply may take.
Schedule
Run an agent on a cron schedule with friendly presets — for digests, checks and proactive outreach.
Workflow triggers
Attach published workflows to when a conversation starts, a message arrives or a conversation closes — blocking the reply or in the background.
Channels
Deploy the same agent to web embeds, Teams, Slack, phone lines and the API without rebuilding it per channel.
Prompt library
Manage prompts like the institutional assets they are.
Keep voice and policy consistent across every agent and AI workflow step. Prompts live in one organised library, with full version history — so you always know what changed, who changed it, and which version produced a given answer.
- Folders and permissions — organise by team and control who can read or edit
- Variables and logic — Liquid templating for names, dates, participant context, conditionals and loops
- Reusable fragments — share brand voice and policy snippets across prompts, including platform-level fragments
- Version history — numbered versions with diffs, line-by-line blame and one-click restore
- Version pinning — pin agents and Generate Text steps to a version; every run records the version it used
14Add deferral escalation ruleStudent services · today 09:12Pinned by 3 agents
13Tone: shorter answersDigital team · 2 days agoRestore
You support students of {{ tenant.name }}.
Today is {{ current_date }}.
+ If a student asks to defer an exam, confirm the
+ deadline and offer to pass them to an adviser.
- Keep answers under 120 words.
+ Keep answers under 80 words.
{% render 'brand-voice' %}
AI Agent Builder
Turn an idea into a working agent — you approve every part.
Describe the problem in plain English, typed or spoken. The builder asks clarifying questions — or uses sensible assumptions — then drafts a complete, editable blueprint. Nothing is created until you approve it.
- A full blueprint — the agent, its workflows, safeguards, assumptions to check and connections to configure
- Knows when not to build an agent — it can propose a workflow-only solution instead
- Runs on your models — pick any approved chat model and set reasoning and verbosity
- Siblings for workflows and code — the AI Workflow Builder drafts processes, and an AI code assistant writes Custom Code steps
We want an agent that helps applicants book open-day sessions and answers questions about entry requirements.
Should bookings be confirmed by email, and should staff approve group bookings over 10 people?
Agent · Open day assistantCourse pages · Gecko Engage events tools
Workflow · Confirm bookingSend message · email
Safeguard · Group booking approvalRequires approval · Teams
Create after reviewEdit blueprint
In conversation
Fast, grounded and careful in every conversation.
Agents stream their replies, call tools in parallel and cite what they used. Controls apply before the model is ever called.
- Streaming replies, including while tools run in parallel
- Citations for every answer — knowledge and web pages with relevance scores and links
- Guardrails first — input is screened before it reaches the model
- Inline identity checks and approvals mid-conversation
- Thread-aware in Slack and Teams, so context carries through a conversation
Has my application been reviewed yet?
I can check that for you. To protect your record, I’ve sent a six-digit code to the email on your application — please enter it below.
482913
Thanks — you’re verified. Your application is under review, and decisions for this programme are usually issued within 15 working days.
admissions-timelines · 0.89engage.findContact
Type your message ↑
Run tracing
RAG you can explain to an auditor.
Every agent run is stored with a full, ordered trace — from the input, through each model call and tool call, to the outcome. Runs link to their conversation and back.
- Model input, output and token usage for each step
- Every tool call — started, succeeded or failed — with its payload
- Website searches show the exact query, each result’s relevance score and the thresholds applied
- Cancel an in-flight run from the console
IN
Initial input“What are the entry requirements for nursing?”
0.0sGR
Guardrailsno rules matched
0.0s⌕
Website search“nursing entry requirements” · 6 results ≥ 0.55
0.4sAI
Model output2,041 in · 264 out tokens
2.2s✓
Run completed2 sources cited
2.3s
Ownership & access
Let departments build — without handing over the keys.
Distributed building only works when access is precise. Gecko lets you delegate to teams while IT keeps the institution-wide view.
Resource owners
Agents record who created them, and list pages filter between “Mine” and “All” — so teams find their own work quickly.
Scoped roles
Custom roles can be limited to specific agents, websites or workflows. Run is separate from write — people can operate an agent they can’t edit.
Audit log
Every create, update and delete is recorded against the person or API key responsible.
Change the AI model behind a live agent from a dropdown. Every answer shows its sources. Every run shows every step.
Keep exploring
Your first agent
From idea to a governed production agent — in under a week, on average.
With managed implementation, Gecko helps define the use case, connect knowledge and tools, set the controls and support go-live.
- A scoped first use case
- The knowledge, tools and guardrails it needs
- A clear measure of success