Every Sparks feature, explained simply — and 3 ways IT teams put them to work
A Spark is Synect's name for a reusable AI agent. At its simplest a Spark is a system prompt with a name. At its most capable it is a multi-step pipeline that searches the web, reads your documents, calls your internal APIs, writes to Slack and Jira, runs on a schedule, and hands work off to other Sparks — all without you writing code.
Because there are a lot of moving parts, the features below are grouped by what they actually do for you: the building blocks, turning one prompt into a workflow, letting a Spark act on the outside world, shipping something you trust, getting it into people's hands, and running it hands-off at scale.
1. The building blocks
Everything a Spark is made of. You can stop right here and have a useful agent.
System prompt, welcome message & starters
The Spark's standing instructions, an optional greeting, and up to six one-tap example prompts shown on the empty screen so users know what to ask.
Model, temperature & quality mode
Pick which of 35+ models runs the Spark — or let Auto choose the best one per message. Set how creative it is, and lock a cost tier (cost-efficient / standard / high-end).
Knowledge base (Bring-Your-Own-Knowledge)
Attach documents (PDF, DOCX, XLSX, text) the Spark always reads. Choose whose files:
owner(your curated docs),user_byo(each user uploads their own), orboth.Capabilities you toggle on
Extra powers granted per Spark — web search, code execution, image generation, live page fetch, org-knowledge search, and more. Only what you switch on is ever available, so the agent stays predictable.
Version history
Every edit is snapshotted automatically. Compare any two versions field-by-field and roll back if a change made things worse.
Sharing & access control
Keep it private, invite specific people (view / edit / owner), publish publicly, or share with your whole org. Optionally let trusted users borrow your API keys so they can run the full Spark.
2. Turning one prompt into a workflow
Pipelines let a Spark do real multi-stage work instead of a single reply.
Multi-step pipelines
Chain several steps, each with its own prompt, model, and tools. Earlier steps feed their output into later ones — research, then draft, then review, then deliver.
Parallel pipelines (DAG)
Independent steps run at the same time in 'waves', so a job that gathers five things at once finishes in roughly the time of the slowest one — often 3–5× faster than doing them in a line.
Pause for user input
A step can stop and ask the user a question, then resume with their answer threaded into the rest of the workflow.
Routing rules
Send certain requests to a different (or cheaper) model based on keywords — e.g. quick lookups go to a fast model, deep analysis to a premium one.
3. Letting a Spark act on the outside world
Four integration paths cover almost anything you'd want to connect.
Spark Actions (any REST API)
Paste an OpenAPI spec and the Spark can call that API mid-conversation. Supports API-key, bearer-token, and full OAuth2 — with automatic token refresh and per-user credentials.
Composio integrations (~105 toolkits)
One-click connectors for Slack, Jira, GitHub, GitLab, HubSpot, Salesforce, Stripe, Google Workspace, Microsoft 365 (Outlook, Teams, SharePoint, OneDrive), Notion, Linear, and ~100 more. Each user connects their own account.
MCP servers
Plug in any Model Context Protocol server: Zapier MCP unlocks 8,000+ apps, or point at your own self-hosted / partner MCP endpoint for richer control.
Office-Aware (Excel + Word)
Read spreadsheets cell-by-cell, trace which cells a formula depends on (and what breaks if you change one), list Word paragraphs, and propose tracked-change edits — real document structure, not a flat text dump.
Image generation in-pipeline
Generate real images inside a step, optionally preserving a reference photo's layout so edits look like the original room or scene rather than a generic stock image.
Integration consent gate
Before any connection is made, the user sees exactly what it can read and write and must tick a box. Enforced on the server too, so a forged request can't skip it.
4. Shipping a Spark you can trust
Tools that stop you from publishing something broken.
Spark Generator (build a Spark by describing it)
Describe what you want in plain English and a meta-Spark builds the prompt, pipeline, capabilities, and integrations for you — then tests and validates it. It refuses to declare a Spark 'done' until validation passes.
Eval-driven testing
Create golden test cases, let an AI judge grade each output against a rubric, and track the pass-rate across versions. A thumbs-down in production can be promoted straight into a new test case.
Image-quality judge (Visual Trust)
For image-producing Sparks, two vision models independently score the actual generated picture on prompt fidelity, text readability, layout, and brand safety — the stricter score wins.
Marketplace publish gate
Optionally block a Spark from going public until its eval pass-rate clears a threshold, and require validation to have just passed before a new version is saved.
5. Getting Sparks into people's hands
How a finished Spark reaches teammates, customers, and the wider community.
Marketplace & discovery
Publish to a searchable public catalog with sorting (popular / trending / newest), a featured strip, tags, and capability badges so people can find the right agent fast.
Ratings, reports & creator profiles
Verified thumbs ratings (shown only after enough votes), abuse reports with an admin moderation queue, and public creator pages so trusted authors stand out.
Fork
Copy any public Spark into your own workspace and customise the prompt or swap a model. The original keeps shipping updates to its own users.
@-mentions & Skills
Type
@spark-namein any chat to use a Spark for a single turn. Stack up to four — one heavy Spark plus lightweight 'Skill' overlays (a brand-voice, a checklist) layered on top.
6. Running hands-off, at scale
Automation primitives for when a Spark needs to run without a human watching.
Batch runs
Run one Spark over up to 10,000 inputs in a single click. Large jobs use the providers' batch APIs at roughly half price; small ones run inline in seconds.
Scheduled Sparks
Fire a Spark on a cron schedule and deliver the result by email, Slack, or webhook — a Monday-morning digest, a nightly report, a weekly summary.
Programmatic invoke API
Trigger a Spark from your own code, CI pipeline, or an alerting tool via a token-authenticated HTTP endpoint. Synect becomes a step in your existing automation.
Cross-Spark delegation
An 'orchestrator' Spark can call specialist Sparks mid-task and combine their answers. Safe by design: max three hops, loops auto-rejected, and all cost rolls up to the person who started it.
Cost caps & guardrails
Set a per-run cost ceiling, auto-disable a schedule after repeated failures, and rely on plan-tier and credit enforcement so spend never runs away.
Three blueprints for IT businesses
The features above are abstract until you wire them together. Here are three patterns IT teams ship today — each one stitches several Spark capabilities into a single agent.
Security-questionnaire & RFP autoresponder
Managed service provider / IT consultancy
- Knowledge base (both mode)
- Office-Aware (Excel)
- Multi-step pipeline
- Batch runs
- Eval-driven testing
- Org sharing
The problem
Every prospect sends a 300-row security questionnaire or RFP in a spreadsheet. Answering them by hand eats two senior days and the answers drift out of sync with your real policies.
The Spark
A Spark in `both` knowledge mode: your SOC 2, ISO 27001, and security policies are owner-attached, while the client's questionnaire is uploaded per conversation. Office-Aware reads the .xlsx row-by-row; a short pipeline drafts each answer grounded in your policies, flags anything it can't substantiate, and a batch run fills hundreds of rows at once at half price.
The payoff
A two-day slog becomes a 20-minute review. Every answer is traceable to a real compliance document, and the eval suite catches it the moment an answer regresses.
Incident-triage & on-call orchestrator
DevOps / SRE team
- Programmatic invoke API
- Cross-Spark delegation
- Parallel pipeline
- Composio (GitHub, Slack, Jira)
- Scheduled daily digest
- Cost caps
The problem
When an alert fires at 3am, the on-call engineer spends the first 15 minutes just gathering context — which deploy went out, what the logs say, who owns the service.
The Spark
Your alerting tool hits the programmatic invoke API the moment a pager fires. An orchestrator Spark delegates in parallel to specialists: one pulls recent GitHub deploys, one summarises the relevant logs, one checks the runbook. It synthesises a root-cause hypothesis, posts it to the incident Slack channel, and opens a Jira ticket — with a per-run cost cap so a noisy night can't blow the budget.
The payoff
The engineer opens their laptop to a drafted root-cause and a ready ticket instead of a blank page. A scheduled version posts a reliability digest every morning.
Tier-1 support & onboarding agent
Internal IT / help desk
- Knowledge base (runbooks)
- Composio (Jira SM, Okta, M365)
- @-mentions + Skills
- Org sharing & roles
- Scheduled digest
- Ratings feedback
The problem
Half the ticket queue is the same handful of requests — VPN resets, access grants, 'how do I…' — and new-starter provisioning is a manual checklist that's easy to get wrong.
The Spark
A shared, org-wide Spark with your runbooks as its knowledge base and Composio connections to Jira Service Management, Okta, and Google Workspace / Microsoft 365 admin. Staff @-mention it with a security-policy Skill overlay so answers stay on-policy. A scheduled run posts a weekly digest of stale tickets and pending offboardings.
The payoff
Repetitive tickets get deflected, provisioning follows the same correct steps every time, and the ratings loop tells you which answers to improve.
Build one in plain English
You never touch code. Describe the agent you want, attach a knowledge base, pick capabilities, and let the Spark Generator validate it before you ship.