AI agents Olas: the biggest onchain AI agent economy runs on SQD
Olas runs the largest economy of autonomous onchain agents. When Polymarket’s infrastructure changed, its subgraphs needed a reindex that would have taken months, if it ever finished. Now, with SQD, it takes a week.
Published
We spent months showing a warning instead of correct numbers, because the only fix was a reindex we couldn’t wait for.
- Onchain agent transactions
- 20.5M+
- Agent-to-agent transactions
- 14.6M
- Daily active agents
- 397
- Faster backfills vs their graph-node estimate ~14M Polygon blocks in one week
- ~10x
Agent economy metrics from olas.network, as of 15 September 2026, 14:00 UTC. Backfill figure from Olas engineering, 3 to 10 September 2026. Before and after
What does SQD power?
Pearl is Olas’ consumer app: an agent app store where people run AI agents they actually own: keys, wallets and results included. Its trading agents, like Polystrat on Polymarket, operate around the clock, onchain, on the user’s behalf. Every trade, position and return is public. The numbers are the product.
Before: broken numbers and a three-month fix
Valory, the core contributor to Olas, tracked Pearl’s prediction-market agents with self-hosted subgraphs, The Graph’s graph-node, on Polygon: markets, wallet creations, agent trade history. Then Polymarket released a series of infrastructure changes, new profile handling, new wallet factories, and the subgraphs fell behind.
Before on self-hosted subgraphs
- ~1,000 trades missing (~5%, growing)
- Full reindex: 2 to 3 months
After on SQD
- Complete trade statistics
- Fixes go out with a deploy
Each significant change meant reindexing. A full backfill on graph-node was estimated at two to three months; the team never completed one. They survived on grafting instead, which still took two to three weeks just to recover a few missing weeks of onchain events. Multiple subgraph versions crawled forward in parallel while the gap grew.
The cost was public. For months, the Olas marketing site showed a warning on Polystrat’s ROI charts: “this metric temporarily shows incorrect values.” The same warning sat in users’ agent UIs. By cutover, the subgraph was missing roughly 1,000 trades, about 5% of the total, and growing. New features waited. Agent engineers worked around missing data.
For a project whose credibility runs on live, verifiable agent metrics, this just would not do. A product outage that could take months to recover from.
The move: Squid SDK + Portal
Olas needed exactly what SQD is built for: performant multi-chain indexing with data that’s cryptographically validated. Correct, not just fast.
The team replaced its self-hosted subgraphs with SQD: Squid SDK for indexing and Portal for data access. The choice fit a principle, too. Olas keeps its stack as decentralised as possible, and Portal serves data from a decentralised lake with cryptographic validation.
When the team went looking for a subgraph replacement, they asked Claude Code for alternatives, and it recommended SQD. “It was easy enough to set up and see the result right away in Portal.” The biggest AI agent economy chose its data layer partly on the recommendation of an AI agent.
Full historical backfill Polygon, ~14M blocks
Full reindex on graph-node
estimated, never completed
Full backfill on SQD
measured, 3 to 10 September 2026
0d30d60d90d
The first full backfill, roughly 14 million Polygon blocks of history, started on September 3 and reached chain head on September 10. One week, measured, against a two-to-three-month estimate they’d never been able to attempt. The indexer now sits at Polygon chain head, lagging by minutes at most.
Portal in production
- Source Polygon and Robinhood Chain Prediction-market agents on Polygon today. The Robinhood Chain protocol deployment is live, with new Pearl use cases to follow.
- Data layer SQD Portal and Squid SDK Squid SDK for indexing, Portal for data access, served from a decentralised lake with cryptographic validation.
- Product Olas agents Pearl, the agent app store, and trading agents like Polystrat on Polymarket, operating around the clock, onchain.
- End user Users and the agent economy People running agents check their results against the chain; people deciding whether to join judge the economy by its published numbers.
What changed
Fixes go out with a deploy
On a subgraph, a bug fix is a reindex. On SQD, it's a deploy. Forward-only changes no longer touch historical data: the processor continues from its cursor, so a fix goes live the day it's written, followed by a simple database migration. When history does need recomputing, a fresh backfill runs on a parallel instance and switches over when it reaches head: about a week, with zero downtime for anything reading the data. Schema fixes that used to queue behind a reindex now just go out.
The numbers are whole again
At cutover the team released a correctness fix that would have required a full reindex under the old stack. Trade statistics are complete and shown to users. No warning label.
New chains on day one
Olas is expanding to Robinhood Chain, a network graph-node doesn't support and SQD does. The protocol deployment is live, with new Pearl use cases targeting Robinhood protocols to follow.
When the data is the product
Most companies lose an internal dashboard when a data pipeline breaks. Olas loses the thing it asks the world to trust. People running agents in Pearl check their results against the chain; people deciding whether to deploy an agent judge the economy by its published numbers. If those numbers are stale, the flywheel stalls.
That’s the standard an agent economy sets for its data layer: not just fast backfills, but the guarantee that correct numbers are never months away.
The numbers
Before and after Source: Olas engineering; freshness measured by SQD, September 2026
- Full historical backfill
- The Graph
- 2 to 3 months estimated, never completed
- On SQD
- ~1 week, measured ~14M Polygon blocks, 3 to 10 September 2026
- Recovering a few weeks of missed events
- The Graph
- 2 to 3 weeks of grafting
- On SQD
- A code deploy forward-only, no reindex
- Missing trades at cutover
- The Graph
- ~1,000 (~5%, growing)
- On SQD
- Complete trade statistics
- Data freshness
- The Graph
- ~4s average lag median ~2s, max ~38s
- On SQD
- ~0.3s average lag median ~0s, max ~4s
| Metric | The Graph (self-hosted) | On SQD |
|---|---|---|
| Full historical backfill | 2 to 3 months estimated, never completed | ~1 week, measured ~14M Polygon blocks, 3 to 10 September 2026 |
| Recovering a few weeks of missed events | 2 to 3 weeks of grafting | A code deploy forward-only, no reindex |
| Missing trades at cutover | ~1,000 (~5%, growing) | Complete trade statistics |
| Data freshness | ~4s average lag median ~2s, max ~38s | ~0.3s average lag median ~0s, max ~4s |
Agent economy served
Source: olas.network, as of 15 September 2026, 14:00 UTC
20,578,364 Onchain agent transactions all chains, all time
14,658,312 Agent-to-agent transactions a subset of the total
397 Daily active agents 7-day average
In their words
Data matters a lot for us: people running agents in Pearl need to see their results correctly, and people deciding whether to join need accurate overall data. When Polymarket changed their infrastructure, we spent months showing a warning instead of correct numbers, because the only fix was a reindex we couldn't wait for. Now fixes go out with a deploy, and a full backfill takes about a week.
Tatiana Priemova Lead Frontend Engineer, Valory More case studies
All case studies →
Running an agent economy, or anything where the numbers are the product? Talk to the SQD team about a data layer that never makes you wear the warning label.