surf-agentic-base

The agentic base layer for SURF: a system of record for what agents do, a referee for claims made from that record, the security primitives an agent needs on a shared platform, and the engineering standard that keeps all of it honest. It is the contracts layer under a set of capability blocks, one per SURF system, that live in their own packages; see docs/architecture/blocks.md.

It is built on the SURF Developer Platform golden path and adopts the common stack wherever the common stack has an answer. It contains only the parts we could not find anywhere else.

The problem it addresses

An agent that works for people on a shared platform acts for someone, touches data of some class, sometimes does things a person should approve, and talks to people who have a right to know it is an AI. Afterwards someone asks what it did: the person it worked for, an auditor, an incident responder, a regulator. Most agents keep a log for their developers and nothing that answers those questions. This service records each run with that account, and keeps it intact.

When a team changes its agent, a new model, a new prompt, a benchmark arm, it also wants to know whether the new one is better, and that comparison is easy to get wrong. The service is a referee for that as well. It does not drive execution: the job orchestrator and the model server are other systems, and the two HPC modules here are a result channel and cluster facts, nothing more.

Adopted, not written here

It does not ship an agent framework, a tracing backend, a metrics store, a dashboard, an experiment tracker, a workflow engine, or an inference server. Every one of those exists and is better than anything we would write. See the reuse ledger for what is adopted and from where.

The record, for any agent

agentic_base.domain.run_record holds one row per run: the transcript the model actually received, the environment it ran in, and who decided its outcome. Every record also says who the run acted for, what class of data it touched and on which isolation tier, whether personal data was redacted before the transcript was written and by what, who approved which action, whether the person was told they were dealing with an AI, and how generated output was marked. Only the tenant and the code revision are required; the rest have defaults, so a writer that does not know yet still writes, and the corpus can tell a run recorded before the answer existed from one recorded after. docs/architecture/cross-cutting.md says which obligation each field serves.

Every write goes into a hash-chained audit log that GET /runs/integrity verifies. A transcript can be redacted on the way in and erased later, by agentic-base-retention under each tenant's retention policy or for one person's request, without breaking that log, and a tenant can export everything it recorded in one request.

An outcome cannot be recorded without naming who decided it, and only some deciders count. A benchmark's own harness or a person whose decision is the reference is citable; the agent's own claim, a quick in-tree check, a user's thumbs-up or another model's score is kept and marked as diagnostic. Experiment trackers that record who decided type it by modality, human, model or code, which cannot tell a user's thumbs-up from a reviewer's decision, and none refuses an outcome that names no one.

Nothing is exported in a private format. agentic_base.provenance turns one record into W3C PROV, an OpenLineage run event and a Process Run Crate, each through that standard's own library, with the scorer and its authority carried as a declared extension whose schema is in docs/schemas/. The same record exports into MLflow as a trace with a feedback assessment; agentic-base-mlflow --tenant <name> sends a tenant's runs from the service's database.

The referee, when you compare

agentic_base.domain.validity adjudicates whether a contrast across arms is sound: benchmark arms, or two versions of a service agent. It detects exclusion channels whose rate differs by arm, which do not cancel in a contrast, and reports the per-arm accounting behind the verdict. item names the unit of work two runs must share to be compared, arm names what is being compared; a record that compares nothing leaves both empty.

The referee is not a new mechanism. Clinical trials have shipped exactly this artifact for two decades: the CONSORT flow diagram, a per-arm accounting of everyone who left the denominator and why, mandatory for publication since 2001, with an extension for AI interventions since 2020. The defect it exposes has a name in the missing-data literature, missingness that is MNAR with respect to the treatment arm. What is missing is not the idea. No experiment tracker implements the check, and we found no reporting standard in agent evaluation that requires the accounting, so validity implements the check and flow_by_arm produces the accounting in the standard's vocabulary: assessed, excluded with reasons, analysed.

The rest of the library

Small things an agent on a shared platform gets wrong when it writes its own: the outbound URL check with DNS pinning, the structural pre-filter over generated code, a completion probe that asks a model for an answer instead of trusting a status code, a retry policy that recycles a dead transport, cluster facts without credentials, a value back from a batch job, the tool contract, the span vocabulary from the standard packages, and the recording seam a served call passes through. Two more that exist because a regime asks for them and the record is where they land: redaction of a transcript before it is written, patterns always and names from a model with a fallback, failing closed rather than claiming more than ran; and a retention policy that refuses to keep less than the law requires, with an erasure the hash chain survives. docs/architecture/boundaries.md lists them; docs/architecture/reuse-ledger.md says for each whether it is adopted, bridged or built, and a test holds the code to that ledger.

Installing

pip install surf-agentic-base                # the library half: four dependencies, Python 3.10+
pip install 'surf-agentic-base[provenance]'  # plus the three provenance-standard libraries
pip install 'surf-agentic-base[service]'     # the service: FastAPI, storage, tracing, MCP
pip install 'surf-agentic-base[mlflow]'      # plus the MLflow export

The import name is agentic_base. The distribution is named surf-agentic-base because agentic-base on PyPI belongs to an unrelated project.

Try it

Three runnable examples, each with its output committed beside it and checked by a test, so what you see below is what you will get.

An agent that works for people. examples/service_agent.py records a merge-request review agent's runs through the client: who each review was for, the data class and tier, how the person was told it was an AI, a maintainer's approval, and three verdicts of different standing. Then it reads back everything the agent did for one person, and checks the records are intact.

pip install 'surf-agentic-base[service]'
export API_TOKENS='{"example-token-0000001": ["example-team", "platform-team"]}'
REDACTION=patterns just run          # in one terminal
python examples/service_agent.py     # in another

1. Three reviews, each recorded for the person who asked
   recorded 3 runs

2. The agent wanted to push to mr-103; a maintainer decided
   approval recorded

3. Verdicts on the reviews, and which of them mean anything
   mr-101  the developer's thumbs-up    user_feedback  diagnostic
   mr-102  another model's score        model_judge    diagnostic
   mr-103  the maintainer's decision    human          authoritative

4. Everything the agent did for alice, from the tenant's export
   mr-101: Automated review: Looks fine; one missing test for the retry path.
     told it was an AI: every review comment begins 'Automated review:'
     data class internal, tier virtualised
     approvals: none
     stored question: Review mr-101. Questions to <EMAIL_ADDRESS>.
   mr-103: Automated review: Suggest a fix: pin the base image by digest. Push it?
     told it was an AI: every review comment begins 'Automated review:'
     data class internal, tier virtualised
     approvals: approved by urn:example:maintainer-carol
     stored question: Review mr-103. Questions to <EMAIL_ADDRESS>.

5. Are the records as they were written?
   intact: 3 runs match 7 entries

Without the service. examples/is_this_comparison_sound.py uses the library half on a record type of its own. Two configurations of an agent attempt twenty tasks; one looks far better only because it times out on the hardest six, and a timeout has no verdict.

pip install surf-agentic-base
python examples/is_this_comparison_sound.py

1. The number people report: resolved, over runs that finished
   baseline      10/19 = 52.6%
   with-planner  12/14 = 85.7%

2. What the validity check says
   not sound: timeout: 30.0% (with-planner) vs 5.0% (baseline)
   examined 40 runs, 2 arms, 1 exclusion channel(s); could have flagged: True

3. The per-arm flow the verdict rests on
   baseline: assessed 20; excluded 1 (1 timeout); analysed 19
   with-planner: assessed 20; excluded 6 (6 timeout); analysed 14

4. Two honest numbers instead of one flattering one
   every run, a timeout counted as unresolved:
   baseline      10/20 = 50.0%
   with-planner  12/20 = 60.0%
   only the 14 tasks both arms finished:
   baseline      10/14 = 71.4%
   with-planner  12/14 = 85.7%

5. Which verdicts may be cited
   the benchmark's own harness        citable: True
   a quick in-tree check              citable: False
   the harness, but it failed open    citable: False
   the agent grading itself           citable: False

With the service. examples/record_and_ask.py records runs from a stand-in agent through the client, including one that crashes, labels them, and asks the service whether the comparison is sound, whether the records are intact, what was kept of a transcript, and for one run in W3C PROV.

pip install 'surf-agentic-base[service,provenance]'
export API_TOKENS='{"example-token-0000001": ["example-team", "platform-team"]}'
REDACTION=patterns just run        # in one terminal
python examples/record_and_ask.py  # in another

1. Recording 12 runs to http://localhost:8080
   the run that crashed was kept: status=failed, failure_kind=RuntimeError

2. Labelling the finished runs with the benchmark's own harness
   a label that names no scorer is refused: HTTP 422

3. Is the comparison sound?
   not sound: timeout: 33.3% (with-planner) vs 0.0% (baseline)
   baseline: assessed 6; excluded 0 (none); analysed 6
   with-planner: assessed 6; excluded 2 (2 timeout); analysed 4

4. Are the records as they were written?
   intact: 12 runs match 22 entries

5. What was stored of the transcript
   (the crashed run has no transcript)
   Fix task-1. Mail the report to <EMAIL_ADDRESS>.
   redaction: agentic-base <version> patterns, no names or places

6. One run in W3C PROV, for a reader that is not this service
   sections: activity, agent, entity, prefix, used, wasAssociatedWith, wasAttributedTo, wasGeneratedBy
   the outcome: {"ab:authority": "authoritative", "ab:degraded": false, "ab:instrument": "harness-1.4", "ab:label_source": "official_harness", "ab:resolved": true}

7. Everything this tenant recorded, to take elsewhere
   manifest says 12 records; the file has 12
   the same request without the token: HTTP 401

<version> is the installed package version.

From a chat client. agentic-base-mcp serves the same database read-only over MCP, with four tools: list_runs, get_run, validity_report and corpus_stats. Point it at the database the service writes, in the mcpServers block of Claude Desktop or a project's .mcp.json. It opens that database itself, so it runs where its user may already read the database and is never published as a network service:

{
  "mcpServers": {
    "agentic-base": {
      "command": "agentic-base-mcp",
      "env": { "DATABASE_URL": "sqlite:////absolute/path/to/agentic-base.db" }
    }
  }
}

Then ask it, for example, whether the comparison for example-team is sound.

get_run returns a long transcript in pages, so no part of a run is out of reach: transcript.next_message says where the next page starts, and a call with from_message set to it returns that page. A page holds at most AP_MCP_MAX_TRANSCRIPT_CHARS characters of messages, 60,000 unless set, and 0 removes the limit. A caller can ask for smaller pages with max_chars. A message longer than the limit is returned whole rather than cut, and the system prompt comes whole on the first page. list_runs returns at most AP_MCP_MAX_ROWS rows, 200 unless set, and says how many there were.

Running locally

just run          # uvicorn with reload on :8080
just check        # tests with coverage, ruff, mypy

SQLite by default so it runs with no infrastructure. PostgreSQL, provided as a platform tenant resource, in every deployed environment.

Every data route requires a bearer token. API_TOKENS maps each token to the tenants it may use, ["*"] for all of them, and comes from the platform's secret management. With no tokens set the service refuses every data request and says what to set; AUTH=none turns the check off for local development, and the service logs a warning when it starts that way.

AUTH=none just run

Documentation

• Blocks and layers: the design, what exists behind each block, the order to build
• Decisions: eleven choices that are cheap now and expensive to reverse
• Reuse ledger: adopt, bridge, or build, with the reason
• Logging, security, safety and compliance: where each lives, the obligations, and what the vision should say
• Observability: what the service emits and where to point it
• What came from agentic-env: what was extracted, what was left, what the first consumer found
• Compliance evidence: what the record produces for the AI Act, NIS2, the Cyber Resilience Act and the Data Act, and what is missing
• Redaction: what removes personal data from a transcript, and what each mode catches and costs
• Incident response: the two reporting clocks, who is told, and what this repository hands you
• Going live on SDP: the deployment repository and what is left to settle

The standard

docs/ENGINEERING.md is why the rules are the rules: what each cost to learn, and the guard that enforces it. CONTRIBUTING.md is the short form for someone about to open a pull request.

Where development happens

Here. Changes land on main through pull requests that the gate has passed on both ends of the supported Python range. SURF's own deployment of this service, the environment overlays and the internal pipeline definition, is kept apart from this repository because it describes where the software runs rather than what it does. The mechanism is general and is documented in docs/architecture/deployment-overlay.md: a deployment repository is this one plus an additive overlay, kept current by merging main, with scripts/overlay.py to compose, check and sync.

Versioning and releases

Semantic versioning, with the version taken from the git tag. While the major version is 0, a minor bump may change the public interface. The import name changed once, from app to agentic_base in 0.2.0, because a top-level app collides with the first package any consumer of a web template already has. Releases are signed vX.Y.Z tags; each is published to PyPI by trusted publishing, with generated notes, build provenance and an SBOM attested on the distribution files. See CONTRIBUTING.md for the release procedure.

Citation

Cite the repository using CITATION.cff; GitHub renders it under Cite this repository.

Licence

European Union Public Licence v. 1.2. See LICENSE.