AI Agent System

01What does the platform accept — a prompt, or a goal?

A user goal with success criteria: the platform decomposes it into a plan, runs a step loop, and keeps a run state the user can watch, pause, and cancel.

02How does the agent touch the outside world?

Only through registered tools with typed schemas — every call is validated before execution and recorded after, so the set of possible side effects is enumerable, never ad hoc.

03What actually enters the model on each step?

A curated context pack, assembled fresh each step: instructions, the goal, the plan, compact summaries of prior steps, and just-relevant memory — never the raw append-everything transcript.

04A step needs human sign-off — what happens to the run?

The run checkpoints its full state and parks; a human approves hours or days later and the run resumes from durable state exactly where it stopped — an approval pause is a checkpoint, not a held process.

05Does the platform remember users across runs?

Yes, as governed memory: facts and preferences carry provenance and expiry and can be corrected or deleted — a stale preference or a revoked permission must not steer a future run.

06How do behavior changes ship?

Through the eval harness: every prompt, tool, or model change re-runs the regression suites — built partly from traced production failures — before rollout, so behavior ships with evidence, not vibes.

01What stops a run from spiraling?

Hard budgets checked before every step — step count, token cost, wall clock — so a stuck loop trips a circuit breaker and halts with a partial-result report, never an unbounded bill.

02A worker dies mid-run — what does the user lose?

Nothing completed is lost: run state persists after every step. External side effects use a stable idempotency key plus an execution record or outbox; a checkpoint alone cannot prevent the sent-email-but-not-checkpointed failure window.

03A run did something odd — can we reconstruct why?

Every step is traced — context pack contents, prompt and model versions, tool inputs and outputs, budget state, the decision taken — replayable end to end.

04Many tenants, shared platform — where are the walls?

Tenants are strictly isolated: memory, tool credentials, and run state are scoped per tenant, and each tool call carries the end user’s own permission, never a platform-wide credential.

05What keeps a long run from getting expensive AND dumb?

Per-run cost is bounded by design: context packs cap per-step tokens so cost grows linearly with steps, and large tool outputs are processed outside the context, not appended into it.