Agents and Command Center
AI agents that do the work while you get on with yours
A roster of Master agents and an Orchestrator that runs them. Describe the job, approve the quote, and watch the plan tick through to a filed deliverable.
Six Master agents, each a real specialist
Each Master has its own tools, default AI model and result layout. Rename it, rewrite its brief, or change its skills, model, engines and app access.
- The OrchestratorThe one to talk to when you don't want to choose: it plans the work, quotes it and runs the specialists.
- The Daily CommanderOpens your day: triages calendar, inbox, reminders and tasks, surfaces what needs you and drafts first replies.
- The Research OracleLive multi-source research that cites primary sources and returns confidence-rated findings in a clean report.
- The DirectorOne brief to a finished multi-scene film, with score, voiceover and stills made in the same run.
- The BuilderWrites working code, runs the tests in a sandbox, then hands you a deploy path or pushes to your GitHub.
- The Music ProducerA complete release from one brief: multi-track songs, lyrics, vocals, stems and cover art.
In detail
Default AI model, and what else each Master does
- The Daily Commander (Claude Sonnet 5.5): reminders, recurring routines, hand-offs to specialists, and send or post permission per run
- The Research Oracle (Claude Opus 5.5): PDF and audio analysis, and a cited report you can download as a typeset PDF
- The Director (Claude Opus 5.5): three creative directions, a locked palette, and up to 3 saved characters cast across shots
- The Builder (Claude Sonnet 5.5): UI mockups, a build plan, and code review with severity-tagged findings
- The Music Producer (Claude Opus 5.5): sequencing, a loudness master and an album zip at no extra cost, and music beds under video
Writing a book? That happens in Novel Mode rather than with these agents.
The Orchestrator: plan first, then execute
Describe a goal in plain English and get back a plan naming which specialist runs each step. Nothing runs until you approve it.
- Describe the goalNo need to know which agent to pick. Choosing is the Orchestrator's job.
- Read the planA plan card shows the goal, the specialist for each step, the order and the handoffs.
- Approve the quoteEach step and the total get a price range before you commit. Edit or fork the plan, then press Execute plan.
- Watch it runEach specialist takes its step with a live card showing status, spend and result.
In detail
What a plan can do
- Steps run in dependency order, each starting once the steps it needs are done
- Re-run a single failed or cancelled step without repeating the rest
- If no Master fits, it can create a new specialist that stays in My Agents; custom agents are capped at 50 per account
- The Council and The Orchestrator are never plan steps
- A plan that cannot work is refused with a plain-English reason
Results you can use, files you can find
Agents deliver real layouts rather than a wall of raw output, file everything they make, and keep the thread open for your follow-ups.
- Media that plays in placeVideo and audio players, image grids, a scene carousel and a full-screen view, each with download.
- Documents ready to downloadReports, build plans and code reviews render as cards, and a deliverable downloads as a typeset PDF.
- Filed without askingEverything lands in Files, under Agents, in a folder named for your request and date; automation and ads runs share a monthly folder.
- Zip a run, or add to oneDownload a run's folder as a zip, or attach files mid-run and the agent is told where they are.
- Follow up with the same agentYour reply goes back to the agent that did the work, which keeps what it already did.
- Run it againOpens a new thread pre-filled with the same brief. Nothing is sent until you press send.
In detail
While it works, and afterwards
- A live milestone stream and task board while a run works
- Each agent keeps a Memory.md that is loaded into every one of its runs
- A run history with View result on any past run
- Per-agent run count, total spend and last-run time, and a flag when a run is waiting on you
- Agent deliverables are kept, not scheduled for deletion
Scene tools on a finished film
- Re-render one scene, billed for that scene only, with the price shown before you confirm
- Re-cut the timeline by reordering or dropping scenes, at no cost
- Add a music bed from your own finished tracks, at no cost
- Build an .srt subtitle file from the film's dialogue, at no cost
Command Center: one tab per job, started from anywhere
Each tab is one background mission with its own conversation and spend meter. Start one from any page, from Chat or from your phone.
- Keep up to 20 jobs open at onceRename, pin, reorder and reopen tabs as you go.
- Type, speak or attachDictate by voice, or attach images, PDF, text, Word, JSON, CSV or Markdown files up to 4 MB.
- Walk away while it worksA mission keeps working in the background, streaming progress into its tab. Cancel at any point.
- Start from any pageThe Agent Bar in the header takes a spoken or typed command.
- Start from ChatAgent Mode turns each chat send into a mission that also appears here.
- Start from your phoneThe Master roster and mission flow, with voice, a spend limit and approvals.
In detail
Tabs and the composer
- Shortcuts: Command or Ctrl + T, W and 1–9 for tabs; Command + A opens the Agent Bar for typing
- Likely steps are shown before you send, on a command of up to 8,000 characters
- Six one-tap starter prompts on an empty tab
- Saved missions: reopen a closed one, or delete it permanently
The Agent Bar, Chat and mobile
- Hover the Agent Bar to see live missions, and set a spend limit before you start one
- A notification when a mission needs input or finishes, linking to its tab
- Agent Mode in Chat keeps its own strip of up to 8 tabs and its own approval gate
- Custom agents launch from desktop; mobile has no picker for them
Build your own agent, and choose the AI model it thinks with
Start from a Master or from scratch, with no code: give an agent a name, a brief, skills, an AI model, render engines and the apps it may use.
- A brief it reads every runName the agent and write the one paragraph it follows every time.
- 127 skills to choose fromEquip up to 24. The picker only offers skills that agent can actually run.
- 15 AI models from 8 providersSet a model per agent, override it per mission, or pick one per follow-up thread.
- App access you grantUp to 40 connected apps per agent, or allow all. Custom agents start with none.
- Your MCP servers as toolsPaste an MCP server URL and its tools become available to your agents, with your approval on first use.
- A brand kit for every runFill it in by hand at no cost, or let AI fill the blanks. Every agent reads it.
In detail
The 15 AI models
- GPT: GPT-4o Mini, GPT-6 Luna, GPT-6.1 Sol, GPT-6 Astra
- Claude: Haiku 4.5, Sonnet 5.5, Opus 5.5, plus Claude Fable 5.1 on mobile
- Gemini: 3.1 Flash Lite, 3.5 Flash, 3.1 Pro
- Grok 4.5, DeepSeek V4 Pro, Qwen 3.7 Max, Kimi K3 and GLM 5.2
The 127 skills, by category
Apps, MCP servers and saving
- Allow-all also covers apps you connect later; built-in Masters keep their access unless you narrow it
- App grants are enforced on every action, while skills steer an agent rather than lock it
- MCP servers: up to 16, with 32 tools per server and 128 in total
- Save a custom agent before it runs, or reset a customised Master to factory defaults
The Council: several models on one hard question
Also on the roster, for decisions where one model is not enough: a panel answers in parallel, then one model writes a single recommendation.
- Pick a panel of 2 to 5From Claude Opus 5.5, Claude Sonnet 5.5, GPT-6 Astra, GPT-6.1 Sol, Gemini 3.1 Pro and Gemini 3.5 Flash.
- Choose who sums upOne model reads every answer and writes the final recommendation.
- See each positionEvery member's answer shows its model, confidence and position, and opens to the full text.
- See shared wordingAn overlap matrix shows wording members share, points only one member raised, and pairwise overlap, with no extra model call.
In detail
How the panel runs
- Default panel: Claude Opus 5.5, GPT-6.1 Sol and Gemini 3.1 Pro, synthesised by Claude Opus 5.5
- The estimate updates live as you change the panel
- Duplicate picks are removed, and it won't convene if a panelist's provider reports an outage
The overlap matrix, shown on desktop, measures shared wording, not shared meaning.
The engines your agents render with
Set a default image and video engine, with resolution, on agents that render, or leave it on Let the agent choose and it picks per brief.
| Engine | Makes | Resolution or quality | Clip length |
|---|---|---|---|
| Kling V3 | Video | 720p, 1080p, 4K | 3–15s |
| Seedance 2.5 | Video | 480p, 720p, 1080p | 4–30s |
| Grok | Video | 720p | 1–15s |
| GPT Image 2 | Image | SD, HD | — |
| Nano Banana Pro | Image | 2K, 4K | — |
| Grok Imagine | Image | — | — |
| Grok Imagine HQ | Image | 1K, 2K | — |
In detail
Consistency and engine pickers
- Cast up to 3 saved characters, so every shot anchors on their reference photos
- Brand palette lock, on The Director, pins your brand kit's colours into every image and scene; it is off by default
- Image engine picker on The Director, The Orchestrator, The Music Producer and custom agents; video engine picker on all of those except The Music Producer
- Naming an engine in your prompt overrides the default for that run
Pay as it works, inside limits you set
No upfront fee on any agent: watch the meter, set a spend limit, and nothing sends without your say.
- No upfront or dispatch feeOn any agent. You pay for the work as it happens.
- Spend you can watchA live spend meter sits on the mission tab and on the agent card.
- A spend limit for any runIf a run would cross it, it pauses. Lift the cap and resume from the checkpoint without re-paying finished work.
- Renders count toward itImage, video and music spend lands on the same limit, and your balance is always the hard stop.
- Waiting on you costs nothingWhen an agent stops to ask a question, nothing is billed while it waits for you.
- Nothing sends without your sayEmail and posts need a per-run grant that is off by default, even when you re-run a past run.
In detail
Sending, posting and hands-off runs
- The send or post grant is offered on The Daily Commander and your custom agents; without it, you get drafts to approve
- The Director, The Music Producer, The Research Oracle and The Orchestrator never publish, so they don't offer it
- Goal Mode runs a Master with no check-ins until your written definition of done is met; it needs its own permission
Pauses and recovery
- The limit slider never goes below what the chosen AI model needs for its first step
- Every pause names its reason in plain English, with the right next step
- A run paused for balance resumes when you top up; if a provider is out of capacity, the run retries without spending
- Retry a step the model returned nothing for, at no cost; cancelling a run also cancels the runs it started
Screening and privacy
- Every brief is screened before work starts, and a blocked brief costs nothing
- Each run gets your profile and this agent's memory of you, with emails, phone numbers and long digit strings redacted; you can preview it
- Up to 20 reference files from the agent's own folder are also considered
- Agent runs are not available in Forget and Burn, the zero-retention privacy mode; switch to Standard or Neural to run one
One account covers every studio: top up as you go, or choose a monthly Pro or Studio plan.
Frequently asked questions
Straight answers about cost, control, models and permissions.
What is an AI agent on SynergyAI?
An agent is an AI specialist you hand a goal to: it plans, does the work, can stop to ask you a question, and files the results. Choose one of six Master agents, including The Orchestrator, which picks the specialists for you, or build your own.
How much does an AI agent run cost?
No agent has an upfront fee: you pay as the work happens. The Orchestrator prices each step and the total as a range before you commit. A run that reaches its spend limit pauses; lift it to resume without re-paying finished work. Top up as you go, or choose a monthly Pro or Studio plan.
What happens when an agent needs my input?
It stops and asks, with 2 to 5 options and its recommendation marked, or room for your own answer. Nothing is billed while it waits, and answering resumes the run where it stopped. Up to 5 questions per run; after 12 hours without a reply it proceeds with its stated default.
Which AI models can my agent think with?
15 models from 8 provider families: GPT-4o Mini, GPT-6 Luna, GPT-6.1 Sol, GPT-6 Astra; Claude Haiku 4.5, Sonnet 5.5 and Opus 5.5; Gemini 3.1 Flash Lite, 3.5 Flash and 3.1 Pro; Grok 4.5; DeepSeek V4 Pro; Qwen 3.7 Max; Kimi K3; GLM 5.2. Mobile also offers Claude Fable 5.1.
Can an AI agent send email or post on my behalf?
Only with a per-run grant you tick yourself. It is off by default and starts off again every time, even on a re-run, and is offered on The Daily Commander and your custom agents. Without it, the agent prepares drafts and posts for your approval.
Can I build my own AI agent without writing code?
Yes. Give it a name and a one-paragraph brief, equip up to 24 of 127 skills, grant the apps it may use, and pick its AI model and default image and video engines. Custom agents launch from My Agents on desktop.
Can an AI agent use my connected apps and MCP servers?
Yes, within permissions you set. Each agent gets its own app grants, up to 40 apps or allow-all, enforced on every action. Runs can also use tools from MCP servers you enable, with your approval on first use: up to 16 servers and 128 tools.
Can an AI agent make a whole video or film?
Yes. The Director takes one brief to a finished multi-scene film with score, voiceover and stills, planned to a target runtime, with one engine rendering every scene: Kling V3, Seedance 2.5 or Grok. Afterwards you can re-render one scene, re-cut the timeline, add a music bed or build subtitles.
Give one job to an agent and see what comes back
Describe the goal, approve the quote, and get a filed deliverable back, inside a spend limit you set.