The Short Answer
BIS economists released MaRCo@BIS on November 12, 2026: an Apache-2.0 framework of 12 plain-text skills, 30 verified model specifications and a tested library that turns an AI coding agent into a supervised research assistant for central bank economists. Its lesson for regulated institutions: fix what can be fixed and gate the rest. On ibl.ai you own all the code and the data, so that pattern runs inside your perimeter.
The release accompanies BIS Working Paper 1395, "MaRCo@BIS: AI-agentic econometric analysis for central bankers", by Marco Lombardi, Taejin Park and Fernando Perez-Cruz of the BIS Monetary and Economic Department. The code is on GitHub at bis-med-it/marco-BIS.
Some context before the details. The BIS is owned by 63 central banks that together account for about 95% of world GDP, and working papers carry the authors' views, not necessarily those of the BIS.
What did the Bank for International Settlements release with MaRCo@BIS?
MaRCo@BIS stands for the MAcroeconometric Research Companion.
According to the working paper, the release holds a tested reference library of 24 estimation modules, 12 workflow skills, a dictionary of 30 canonical model specifications verified against the original papers, and a data layer with runtime connectors.
The library spans VARs, Bayesian VARs and local projections, with recursive, long-run, sign, narrative, proxy and heteroskedasticity-based identification.
Pre-collected panels include the FRED-MD and FRED-QD datasets. The repository README says the reference library is covered by more than 1,000 tests.
The 12 skills listed in the repository are knowledge-base, orchestrator, model-selection, model-dictionary, specification, data-retrieval, diagnostics, results-presentation, benchmarking, consistency-audit, advisor-protocol and research-log.
The package is distributed under the Apache-2.0 licence, without warranty.
How does the BIS framework stop an AI agent from making silent decisions?
The paper's first design principle is a guided, research-led workflow.
At every consequential decision, such as the model class, the variable set, the identification scheme, the sample or a data substitution, the agent proposes a choice with its rationale and waits for the economist to confirm or override it.
A session starts in one of three modes: replicate a paper already in the dictionary (mode A), ingest a new paper and verify its specification (mode B), or design a specification for an open question (mode C).
All three run the same staged loop of specification, data, estimation, an economic sensibility check, reporting and a final disposition.
Two rules govern specification. Variables and identification are never chosen by statistical fit, which settles only tuning choices like lag order.
And specification search is disclosed: the final report counts sensibility-driven revisions and shows first-pass results next to final ones.
Every decision enters a research log with its rationale, the alternatives considered and the human confirmation. Four days later, Federal Reserve Governor Michael Barr asked the same of AI decisions in banks: that they be explainable and replicable.
What changed when the BIS ran the same request with and without MaRCo@BIS?
The authors ran paired sessions of two replication exercises on the same day and the same underlying model: once with the framework, once in a plain coding-agent session in an empty folder. Their Table 4 summarises the difference, condensed here.
| Dimension | With MaRCo@BIS | Plain coding agent |
|---|---|---|
| Decisions put to the researcher | Gates at the binding moments: data source, substitutions, disposition of new code | Deep design calls made autonomously |
| Data sourcing | Known sources, confirmed by the researcher | Web search at run time |
| Estimation code | Shared library backed by the test suite | Several hundred lines of fresh, untested code per run |
| Validation | Machine-checked targets from the dictionary entry | Self-asserted match to the paper's figures |
| Diagnostics | Sensibility gate plus residual diagnostics | None run in either session |
| Record left behind | Research log, run state, data registry, disposition | Clean code and a readme; the decision trail lives only in the chat |
Both setups reproduced the headline finding. The difference was traceability. In one plain session the agent estimated the VAR with a constant term the original paper explicitly excluded, an error that leaves the headline intact and is easy to miss.
The framework did not come free. The paper reports that on the first exercise both sessions used about 1.6 million weighted tokens; on the second, the framework session used about 3.7 million against about 2.2 million without it.
The authors say two short sessions cannot settle which effect dominates.
How well did the BIS AI agent replicate published macroeconomic models?
The authors benchmarked the agent's replications of three widely used structural VARs against the originals and ran five exercises, from a verified replication to an open design question.
The BIS summary states it reproduced published findings where the specification and data were well defined.
It also highlighted the greater uncertainty that comes with set identification and incomplete data.
In the open fiscal policy exercise, the impact multiplier stayed roughly between 0.6 and 1.1 across most lag and shrinkage checks (a one-lag version gave 0.43), while the eight-quarter cumulative multiplier swung from -0.56 to +1.04 depending on the specification.
The agent recommended headlining the impact result and reporting the rest as unidentified.
When the authors chose a Bayesian specification it had argued against, it built the run as instructed and wrote its disagreement into the output. The BIS summary of the paper concludes that AI should support, rather than replace, expert judgment.
Where does central bank data go when an AI research agent reads it?
This is the question a risk officer should ask first, and the MaRCo@BIS README answers it plainly. Agent sessions send prompts and any file contents the agent reads, including data files, to the model provider's API, which for Claude Code is Anthropic's.
The README tells users to clear confidential or unpublished data with their institution's policy first, including which deployment route and data-retention terms it has approved.
An optional advisor protocol can send math questions to a second model; the README says it is meant to carry equations, never data, and calls that an instruction to the agent rather than a technical filter.
That is honest documentation, and it puts the decision where it belongs. For public FRED data the question is easy. For supervisory data, unpublished forecasts or a bank's own positions, the route the data takes is the whole question.
Why does MaRCo@BIS store its knowledge in plain-text files instead of a model?
The paper's fifth design principle divides labour between the stochastic and the deterministic. The AI handles judgement.
Canonical specifications live in the dictionary, estimation code in the tested library, and the conduct of each stage in written instructions the agent follows.
The authors report fewer errors once tasks moved out of the model into fixed components, because errors there could then be corrected centrally.
The same structure makes the framework more general across AI models: there is no server and no bespoke model, so an upgrade or replacement calls for re-validation, probably without a complete rebuild.
When their own model was updated during development, the instructions, specifications and code stayed usable.
The paper notes the work was built on Claude Code running Claude Opus 4.8 and Fable 5 between May and July 2026. Skills that outlive the model are the same idea we covered in agent skill catalogs as a supply chain.
How can a bank or central bank run a MaRCo@BIS-style agent on infrastructure it owns?
The pattern has five parts: plain-text skills, verified reference knowledge, confirmation gates, a decision log and model portability. None of the first four depends on where the model runs. What depends on hosting is the data path the README warns about, and who holds the logs.
On ibl.ai you own all the code and the data.
Institutions deploy the platform in their own cloud, on-premise or fully air-gapped, run it model-agnostic across commercial or open-weight LLMs, and keep AI agents, their instructions and their logs on infrastructure they run. It all runs with no per-seat pricing.
| MaRCo@BIS principle | The question for a regulated deployment |
|---|---|
| 12 plain-text skills | Who can change an agent's instructions, and is the change versioned? |
| 30 verified specifications | Where does the reference knowledge live, and who curates it? |
| Confirmation gates | Which actions can an agent take without a named human approving them? |
| Research log | Does the decision trail stay on systems the institution controls? |
| Model portability | Can the institution switch models and re-validate without rebuilding? |
ibl.ai is family-owned and operated from New York, NY. For banks and supervisors weighing these questions, our financial services page covers deployment options, and Agentic OS is where the agents, their permissions and their logs run.
What does MaRCo@BIS not prove about AI agents in central banking?
It is a proof of concept, in the paper's own words, and the repository states it is not maintained, with issues and pull requests likely to remain unresolved. The value is the published design and benchmark, not a supported product.
The authors list the risks themselves. A user can bypass the safeguards against specification search by starting a fresh session and reporting only the chosen result.
Automation bias remains, so the framework is meant for trained economists, and outputs required expert review that, in their exercises, materially changed the specifications.
The README adds that verification is best-effort and that many methods rest on the library's own simulated-data tests rather than independent implementations.
Read MaRCo@BIS as evidence that disciplined agent design works in a conservative institution, not as evidence that the judgement can be delegated.