The thinking behind the swarm.
Ten notes on why high-stakes work should be run by thousands of narrow agents behind one deterministic boundary — from the mathematics of compounding error to what ten sessions on real capital taught us. Point at a note and the figure becomes its idea.
-
01
Research note
Compounding Error and the Case for Decomposition. A Research Note on Agent Architecture
Why long-horizon autonomous work must be decomposed into narrow agents behind a deterministic boundary.
-
02
Research note
Specializing a Model to a Rulebook and Proving What It Did. A Research Note on Post-Training and Audit
How a general model comes to follow one rulebook, and how a hash-chained ledger proves what it did.
From the unit of work to the open platform.
Read in order, the eight posts walk the architecture from the lifecycle test to the master-and-copy model — each one describing the research idea behind a layer of the stack.
-
03
Series · 01
Not Tasks. Lifecycles. Why We Stopped Building Assistants
Why we stopped building assistants and changed the unit of work to the lifecycle.
-
04
Series · 02
Why Five Thousand Small Agents Beat One Large One
Failure analysis, not a taste for scale: why thousands of narrow agents beat one large one.
-
05
Series · 03
Agents Decide. Code Executes. The One Boundary That Matters
Eight layers, one boundary. The top four estimate; the bottom four are deterministic code.
-
06
Series · 04
Post-Training a Model on a Rulebook, Not on the Internet
Closing the gap between a model that knows finance and one that knows this lifecycle.
-
07
Series · 05
The Execution Council. Governance in Milliseconds
Governance in milliseconds: what the council checks, three operating modes, and the kill switch.
-
08
Series · 06
A Ledger Nobody Can Edit. Hash Chains as Audit Infrastructure
Append-only, hash-chained, replayable. Why a record an administrator can edit is only a claim.
-
09
Series · 07
Lifecycle 01. What Ten Sessions on Real Markets Taught Us
What ten live sessions on real capital showed, precisely what they did not, and where the cost goes.
-
10
Series · 08
Master and Copy. How a Proven Agent Becomes a Platform
How a proven agent becomes a platform: prove one, master and copy, then open the rails.
Before the swarm, the research.
Research summaries, market-structure economics and the first notes on AI agents in finance, from the people behind OpenEXA and the researchers they work with.
More on financial markets researchAI & agents
The first notes on AI copilots, guardrails and autonomous agents in financial markets — the thinking that became OpenEXA's agentic infrastructure.
BrowseFinancial markets research
Summaries of peer-reviewed research in computational finance: ETF price dynamics and drawdown risk, optimal execution, futures portfolios, sparse mean-reverting portfolios and multiscale signal processing.
BrowseMarket structure & economics
Summaries of academic research on how financial markets shape real decisions — information, contracting, credit, automation and the labour share.
BrowseThe research contribution is four constants.
Agents decide. Code executes.
The top four layers estimate and reason; the bottom four are deterministic. The boundary between them is a typed schema.
No single agent sees the whole job.
Each agent is scoped to one gate and does one thing. Its failure is local and its output is a proposal until something downstream permits it.
Nothing an agent believes can move what it is not permitted to.
Capabilities are held per agent and per tool, enforced by the runtime rather than by the model's good behaviour.
Anything that settles is written to a ledger nobody can edit.
Append-only, hash-chained, replayable — and written only after an independent counterparty confirms.
The ideas, as working objects.
Tighten the council on the live swarm. Edit a record and watch the chain break. Run one simulated session of Lifecycle 01.