Runtime Profiles

Runtime profiles are the fastest way to shape how an autonomous agent behaves without tuning dozens of knobs by hand.

Why profiles exist

Each built-in profile bundles tradeoffs across:

• tool call budget
• context budget
• summarization thresholds
• memory read/write policy
• delegation depth

That makes profiles operational presets, not just marketing labels.

Built-in profiles

Profile	Best for	Tool budget	Parallel tools	Context budget	Delegation
fast	Short tasks, fast UI turnarounds, mostly direct answers	8	3	32000	off
balanced	Default product integrations and general autonomous workflows	20	5	96000	role-based
deep	Heavier investigation, larger context windows, more tool-heavy work	40	8	200000	role-based
research	Long-running research agents with broad context and stronger delegation	80	10	400000	automatic

Profile defaults assume a modern frontier model (Claude 4.x, GPT-4o, Gemini 2.x) with at least a 128k context window. If you are running an older or smaller model, pass an explicit limits.maxContextTokens to clamp the budget.

Practical differences

Profile	Last turns kept	Child context policy	Memory scope	Summarization trigger
fast	8	minimal	session	24000
balanced	16	scoped	session	72000
deep	30	scoped	workspace	150000
research	60	full	workspace	300000

Important default to remember

All built-in profiles currently default planning to off.

That is intentional. Autonomous planning is powerful, but it should be enabled because your workflow needs it, not because the profile silently turned it on.

Recommended starting point

Start here unless you know otherwise:

const agent = createSmartAgent({
  model,
  tools,
  runtimeProfile: "balanced",
  planning: { mode: "todo" },
});

Use that as the baseline for autonomous product work. Move up to deep or research only when you can explain why the extra context and delegation are needed.

Custom profiles

If a built-in profile is close but not right, extend it:

const agent = createSmartAgent({
  model,
  tools,
  runtimeProfile: "custom",
  customProfile: {
    extends: "balanced",
    limits: { maxToolCalls: 10, maxContextTokens: 18000 },
    planning: { mode: "todo" },
    context: { lastTurnsToKeep: 10 },
  },
});

Use custom profiles to tune behavior, not to discard the built-in operating assumptions without reason.

How to choose well

• Choose fast if latency matters more than broad exploration.
• Choose balanced if the agent is a product feature and not a research sandbox.
• Choose deep if the agent must inspect more material before it acts.
• Choose research if the agent is explicitly allowed to branch, search, and delegate more aggressively.

Common mistake

The common failure mode is starting with custom too early. You lose the benefit of tested presets and end up debugging self-inflicted configuration drift.