I got very used to running agents locally.

The workflow was simple: run the agent, let it write outputs into my filesystem, then inspect everything in an ./outputs folder.

Markdown reports, JSON files, screenshots, charts — whatever the agent produced, it was right there.

Then I deployed it.

Same agent, same logic. But now the "output" lived in a container filesystem that vanished the second the task finished. A retry wrote report_20260313_103042.pdf next to report_20260313_103041.pdf. And when I wanted to share this with someone, I no longer had a clean link.

Nothing about the agent had changed.

Everything about the environment had.

If you build agents that produce files (reports, datasets, images, JSON dumps), you've probably hit this gap.

Local development hides it.

Production hands it to you on day one.

The local version

On your machine, persisting agent output is trivial:

from pathlib import Path

def save_report(content: bytes, run_id: str) -> Path:
    out_dir = Path("./outputs") / run_id
    out_dir.mkdir(parents=True, exist_ok=True)
    path = out_dir / "summary.pdf"
    path.write_bytes(content)
    return path

That's it. Write bytes, get a path, move on.

You can list the directory, cat the file, open the PDF, or hand the path to the next script in your pipeline.

For one person running one agent on one laptop, this is perfectly fine.

The problem is not local development.

The problem is mistaking “it works on my laptop” for “I have a storage layer.”

The production version

Production agents do not get a reliable ./outputs/ folder.

They run in environments where the filesystem is temporary, isolated, or both.

Serverless functions may give you /tmp, but it is scoped to the execution environment and often limited in size. Containers lose local state when they restart. Background workers, queues, and orchestrators can run each task on a different machine.

And retries are not an edge case. They are part of the system.

Your orchestrator will eventually rerun a failed step, and now you have the same logical output produced twice.

Then there is the human in the loop.

Agents produce things people actually need to read: compliance PDFs, analysis summaries, generated slides, CSV exports, charts, screenshots, debug bundles.

Those people do not have SSH access to your worker node.

They need a link, not a filepath on a machine they will never see.

So the production checklist starts looking very different from local dev:

I have talked to a few teams that hit the same wall.

The agent logic is done.

Now the artifact plumbing begins.

Files vs artifacts

Here's the distinction that changed how I think about this:

A file is bytes at a path.
An artifact is a file plus context.

That context is what makes the output usable after the agent is done.

For example:

• which run produced it: session_id, pipeline run, batch job
• which agent produced it: agent_id, stage, model version
• what it actually is: content type, size, custom metadata
• when it should expire: TTL or lifecycle rules
• how to retrieve it later without knowing where the bytes live: a stable ID
• how to share it with someone outside your infra: a temporary download link

A PDF sitting on disk is a file.

A PDF tagged with session_id=pipeline_run_42, agent_id=report-writer, model=claude-sonnet-4, retrievable as art_2xk9f7v3m1p0, and set to expire in 30 days?

That is an artifact.

Your agent may still produce files.

But downstream agents, debug tools, production workflows, and the humans waiting in Slack all need artifacts.

The metadata layer you eventually build

Most teams do not start by building an artifact store. They start with S3 (or R2, or GCS) and a slowly growing feeling that object keys aren't enough.

The pattern I keep seeing, including in our own user research, goes like this.

First, put the bytes in object storage:

import hashlib
import boto3

s3 = boto3.client("s3")
BUCKET = "my-agent-outputs"

def upload_file(local_path: str, tenant_id: str) -> str:
    data = open(local_path, "rb").read()
    content_hash = hashlib.sha256(data).hexdigest()
    ext = local_path.rsplit(".", 1)[-1]
    key = f"{tenant_id}/{content_hash}/{ext}"
    s3.put_object(Bucket=BUCKET, Key=key, Body=data)
    return key

Then you realize the object key is not enough.

You need to know which run produced the file, which agent created it, what kind of output it is, when it should expire, and how to find it later.

So you add a metadata table:

CREATE TABLE artifacts (
  id            text PRIMARY KEY,
  tenant_id     uuid NOT NULL,
  filename      text NOT NULL,
  content_type  text NOT NULL,
  size_bytes    bigint NOT NULL,
  content_hash  text NOT NULL,
  session_id    text,
  agent_id      text,
  metadata      jsonb NOT NULL DEFAULT '{}',
  expires_at    timestamptz NOT NULL,
  created_at    timestamptz NOT NULL DEFAULT now(),
  deleted_at    timestamptz
);

CREATE INDEX idx_artifacts_session
  ON artifacts (tenant_id, session_id, created_at DESC)
  WHERE deleted_at IS NULL;

Then wrap it in an API:

def create_artifact(file_path, session_id, agent_id, metadata=None):
    key = upload_file(file_path, tenant_id=current_tenant())
    artifact_id = f"art_{generate_id()}"
    db.execute(
        """
        INSERT INTO artifacts
          (id, tenant_id, filename, content_hash, session_id, agent_id, metadata, expires_at)
        VALUES (%s, %s, %s, %s, %s, %s, %s, now() + interval '30 days')
        """,
        (artifact_id, key, session_id, agent_id, metadata or {}),
    )
    return artifact_id

Congratulations, you're on your way to building an artifact store.

Then the other 80% shows up:

• Idempotency keys for retry-safe uploads
• Content-hash deduplication so retries don't double your storage bill
• Presigned upload/download URLs with correct expiry semantics
• Session sealing so a zombie sub-agent can't append to a "finished" run
• Public download links for humans, with their own expiry model
• Soft delete plus garbage collection
• Usage metering and quota enforcement
• Metadata filtering that's actually queryable

I've watched engineers spend time building this type of wrapper and still not get dedup, TTL, or session semantics right.

This is not a knock on those teams. It is necessary plumbing. But necessary plumbing is still plumbing - and most teams should be spending that time on their product, not rebuilding agent infrastructure.

What an artifact layer should handle

If you are deciding whether to build this yourself or use a purpose-built layer, this is the basic checklist I would use.

Run / session grouping

You need to answer one question quickly:

What did this pipeline run produce?

Not grep logs.
Not list an S3 prefix and hope the naming convention held.

One query:

artifacta ls --session pipeline_run_42

A session should be whatever your orchestrator already uses: pipeline_run_42, daily_batch_20260313, customer_report_8841.

It should not require a separate “create session” step just to group outputs.

Agent provenance

When a report looks wrong three weeks later, you need to know what produced it.

Which agent?
Which model?
Which stage of the workflow?

That means agent_id and metadata should be captured at upload time, not buried in logs you hope still exist.

client.push(
    "analysis.json",
    session_id="pipeline_run_42",
    agent_id="summarizer",
    metadata={"model": "claude-sonnet-4", "stage": "final"},
)

Metadata

Object storage metadata is not enough.

Headers are limited, awkward to query, and easy to make inconsistent across a pipeline.

You want structured metadata stored with the artifact record and filterable when listing artifacts.

Deduplication

Agent systems usually need two forms of deduplication:

1. Content-hash deduplication: same bytes, one stored blob.
2. Idempotency keys: same retried operation, same response.

These solve different problems.

Content hashing prevents duplicate storage. Idempotency prevents a retry from creating a second logical artifact.

Conflating the two is a common bug in homegrown wrappers.

TTL

Artifacts should expire by default.

An experiment, batch run, or debug file should not live forever because nobody remembered to clean it up.

Storage lifecycle rules help, but they usually operate at the bucket or prefix level. They do not understand your artifact metadata, which makes per-artifact expiration harder than it should be.

Download links

Humans need a link, not a file path.

A good artifact layer should make it easy to create a stable download URL with configurable expiry:

https://dl.example.com/lnk_...

That link should be separate from your internal storage details and easy to share with a teammate, customer, or workflow step.

Retrieval by ID / session

Downstream agents should not coordinate through shared filesystem paths.

Agent A pushes an artifact and gets an ID. Agent B pulls by ID, or lists the session and filters by metadata.

export ARTIFACTA_SESSION_ID="pipeline_run_42"

python extract.py    # pushes CSV
python analyze.py     # lists session, pulls CSV, pushes report
python notify.py      # creates download link for the human

Session sealing matters too.

Once a run is finalized, late uploads should fail clearly instead of silently corrupting the run:

409 Session 'pipeline_run_42' is sealed. No new artifacts can be added.

A small tool I'm building for this

I'm building Artifacta, an artifact store purpose-built for AI agents.

It is not an orchestrator, search engine, or agent framework. It is the layer between your agent and object storage: session-aware, queryable artifact storage with a CLI, MCP, Python SDK, and REST API.

For example:

pip install artifacta-cli
export ARTIFACTA_API_KEY="ak_live_..."

artifacta push report.pdf --session earnings-q4-2025 --agent report-writer
artifacta ls --session earnings-q4-2025
artifacta link art_2xk9f7v3m1p0   # share with a human

Or from Python:

from artifacta import Client

client = Client()
artifact = client.push("report.pdf", session_id="earnings-q4-2025")
print(artifact.id)  # art_2xk9f7v3m1p0

I’m sharing it because this is a problem I keep seeing in agent workflows, even if Artifacta is not the solution every team chooses.

Questions for other builders

I’m curious how other teams handle this today:

1. Where do your agent outputs go in production?
2. How do you group outputs by run?
3. What broke first: retries, deduplication, human sharing, or cleanup?
4. Do you finalize runs, or can late workers still write outputs?
5. Is a dedicated artifact layer useful, or is “S3 + Postgres + a thin wrapper” the right tradeoff?

Drop your setup in the comments. I’m especially interested in approaches that are not just object storage plus glue code.