Intelligent Iterations

Agent Runtime

Bring your harness.
Keep your infrastructure.

Thanks for joining!
ii-agent-runtime · developer preview

$ ii-agent-runtime deploy aws codex-production

setup rev 13 · sha256:7c9e41b2 · codex 0.42.0 · gpt-5-codex · repo@5034eb3

access 5 credential refs checked live · 0 escalation paths · prod unreachable

launch alice@acme.com authorized by your identity provider

infra runs in aws

✓ deployed codex-production rev 13

$ ii-agent-runtime run codex-production --task sympy__sympy-20590

✓ run-043 42.8s · 182.4k in · 6.9k out · $0.13 · trace, I/O and outputs kept

✓ evaluated SWE-bench Verified with your evaluator: resolved · rerun from rev 13

01 / Secure

Prevent agent privilege escalation.

Configure GitHub, AWS, Google Cloud, Azure, or custom credentials and see everything your harness can and can't do.

permission check · codex-production rev 12

live check · analyzed as ii-analyzer (read-only) · target credentials not elevated

refs github-app/ii-runtime-prod · arn:aws:iam::123456789012:role/ii-agent

ii-agent@acme-prod.iam.gserviceaccount.com · userAssignedIdentities/ii-agent

vault://secret/ii/billing-api

can 47 actions contents:write pull_requests:write s3:GetObject ...

cannot 12,406 actions secrets:read iam:CreateAccessKey roleAssignments/write ...

BLOCK escalation aws iam:PassRole + ec2:RunInstances -> role/deploy-admin

BLOCK indirect github contents:write can push a branch that runs deploy.yml with PROD_TOKEN

BLOCK regression gcp storage.objects.delete reachable since rev 11

WARN excessive aws s3:DeleteBucket granted, never needed by this setup

WARN redundant azure Storage Blob Data Reader already covered by Blob Data Contributor

PASS assertion must not reach Key Vault kv-prod

PASS custom test billing-api refunds endpoint unreachable

deployment gated: 3 blocking findings

02 / Deploy

Make every run reproducible.

Every run stays tied to the exact setup that produced it, so you can see what changed and recreate it.

deploy.json

{

"setup": "codex-production",

"revision": 12,

"harness": "@openai/codex@0.42.0",

"workspace": {

"repository": "intelligent-iterations/ii-agent-runtime",

"commit": "5034eb3dfe3829edb3282d327592735c7b1193b2"

},

"runtime": {

"image": "ghcr.io/intelligent-iterations/agent-base",

"digest": "sha256:1d4028a6ede6bebe8f49c3ab039b22b8a2e785f854e25acaa7ca9073889cc758",

"cpu": "4",

"memory": "8Gi",

"region": "us-east-1"

}

}

03 / Measure

Turn runs into evidence.

Capture the full execution so you can trace what happened, measure what it cost, and evaluate the result with SWE-bench, Harbor, or your own evaluator.

run-042 · execution record

trace 19518d6f1a9d2edf531e4c8df9f69fd8 · run-042 · sympy__sympy-20590

invoke_agent codex 42.8s in 182.4k (151.0k cached) out 6.9k $0.13

chat gpt-5-codex 9.6s

execute_tool shell 1.1s grep -rn "__slots__" sympy/core

execute_tool apply_patch 0.4s sympy/core/_print_helpers.py +5

execute_tool shell 18.2s pytest sympy/core/tests/test_basic.py exit 0

exported run-042.otlp.json · predictions.json

$ swebench eval verified -p predictions.json --run-id run-042

Instances resolved: 1

Make it part of your project

How teams run coding agents in production

Keep your models, identities, credentials, compute, and telemetry storage. We’re building the reusable infrastructure around them. Use the pieces you need; extend the rest.

Thanks for joining!