85 agentic tools audited for security incidents · Talk to the lab →
MotionticBook a call
Glossary

What are artifacts in AI agents?

In AI agents, artifacts are named objects kept outside the prompt, either large data the agent loads by handle only when needed, or reviewable outputs such as plans, screenshots and recordings that let people check the agent's work.

The word is used in two related ways, both from Google in late 2025. In its Agent Development Kit, an artifact is a large object, such as a file or dataset, stored outside the conversation and referenced by name. In its Antigravity development platform, artifacts are the plans, task lists, screenshots and recordings an agent produces so a person can verify its work.

What the two share is a refusal to put bulk into the transcript. One keeps raw data out of the model's context; the other keeps raw tool logs out of the reviewer's way.

Artifacts as externalised data

  • Named and versioned

    Google's context-aware framework post (2025-12-04) describes artifacts as named, versioned objects managed by an artifact service. Its example is a 5MB CSV that should never be pasted into chat history.

  • The handle pattern

    By default the agent sees only a lightweight reference: the name and a short summary.

  • Ephemeral expansion

    A load tool brings the artifact's content into the working context for the call that needs it, and it is offloaded again afterwards by default.

Artifacts as reviewable output

  • What they are

    Google's Antigravity announcement (2025-11-20) lists task lists, implementation plans, walkthroughs, screenshots and browser recordings.

  • Why

    Reading raw tool calls is tedious. The post's heading puts it as verify with artifacts, not logs: a reviewer checks the agent's reasoning at a glance.

  • Feedback in place

    People comment on an artifact the way they would on a shared document, and the agent takes the feedback in without stopping its run.

Files between agents

Anthropic's March 2026 harness post uses files the same way between agents rather than between agent and person. The generator and evaluator communicated through files: one wrote a sprint contract proposing what it would build, the other reviewed it and replied in the file or a new one.

Where it fails

  • A misleading handle

    If the summary on a data artifact is vague, the agent never loads it, or loads the wrong one.

  • Polish is not proof

    A neat plan or a passing screenshot shows what the agent chose to present. Tests and checks behind the artifact still decide whether the work is correct.

  • Version confusion

    When artifacts are revised, the agent and the reviewer need to be looking at the same version.

How Motiontic's engineers apply it

Motiontic's engineers keep large inputs, such as documents and exports, out of the prompt behind named handles, and have agents produce a short plan and a result summary at fixed points for review. Each review artifact links to the evidence behind it, such as test output or the source records.

The two meanings of artifacts in AI agents

The two meanings of artifacts in AI agents
Data artifactsReview artifacts
Origin of the usageGoogle ADK context guidanceGoogle Antigravity
AudienceThe modelA person
ExamplesFiles, datasets, large tool outputsPlans, task lists, screenshots, recordings
PurposeKeep bulk out of the context windowMake the agent's work quick to verify
Main riskA handle that never gets openedMistaking presentation for verification

Common questions

What are artifacts in AI agents?

Named objects kept outside the conversation. Either large data the agent loads by reference only when needed, or outputs such as plans and screenshots that let people review what the agent did.

Why use artifacts instead of putting data in the prompt?

Large data in the prompt costs tokens on every call and dilutes the model's attention. A handle costs a few tokens, and the content loads only for the step that needs it.

Do review artifacts replace testing an agent's work?

No. They make the work faster to check, but a plan or screenshot reflects what the agent chose to show. Tests and validation still decide whether the result is correct.

Running agents that have to hold up?

Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.