01 / Map
Control
Authority, policy, tools, data, mutations, retry semantics, recovery states, and trust boundaries.
Fixed-scope engagement · 5 business days · $5,000
Agent Reliability
& Control Sprint
For CTOs and engineering leads responsible for one production or near-production tool-using agent. In 5 business days, AXIOM turns unclear authority, external-effect, retry/recovery, and runtime-identity behavior into executable tests, one prioritized hardening patch, and evidence your team can rerun.
Scope
One workflow. Explicit boundaries.
The $5,000 fixed price covers one primary repository or service, up to five tool or external-effect types, up to two model/provider profiles, and at most one retrieval subsystem. The sprint is for a system that already exists; it is not a broad rewrite or greenfield agent build.
02 / Test
Failure
Executable deterministic tests for policy rejection, effect safety, ambiguous outcomes, evidence identity, and applicable retrieval boundaries.
03 / Harden
One slice
One prioritized control patch or executable reference patch, bounded to the agreed workflow and backed by a regression test.
Deliverables
A bounded control upgrade your team can rerun.
You leave with the failure boundaries made explicit, one highest-value control hardened, and a reproducible evidence pack that shows what passed, what remains unsupported, and what to do next.
Control map + acceptance packBoundaries and testsIn-scope effects classified; deterministic failure and acceptance cases captured as executable evidence.↓
Bounded hardening sliceImplementationOne prioritized control patch or reference patch with a reproducible regression test.↓
Evidence pack + remediation queueHandoffCommands, revisions/profile identifiers, results, limitations, and ranked remaining work, followed by a 60-minute technical review.↓
Acceptance
Pass or fail is defined before handoff.
The sprint is accepted only when every applicable criterion below passes. Unsupported guarantees remain explicit.
- 01
Pre-effect rejection
Policy-invalid or schema-invalid cases produce no in-scope external effect.
- 02
Fail-closed recovery
Ambiguous non-retry-safe outcomes do not authorize automatic replay.
- 03
Reproducible evidence
One documented command reruns the acceptance pack; failures exit nonzero and tested identity is machine-readable.
Technical proof
Built from validated AXIOM control and recovery work.
Current evidence includes 159 passing APEX offline tests, 18/18 ASON-to-APEX integration tests, and 63/63 ASON suite tests. APEX revision 67696d6 with Gemini 3.8 Flash and google-genai 1.66.0 passed the live provider smoke, 2/2 integration tests, direct execution, and 12/12 benchmark tasks under its recorded execution-profile digest. AXIOM RAG implements explicit storage ownership, replacement recovery, embedding-space identity, and fail-closed HTTP boundaries. Acceptance remains bound to the recorded source, provider, model, SDK, configuration profile, and workload; materially changed profiles require separate validation.
Fit
Start with one consequential workflow.
Good fit: one consequential agent workflow already calling tools, APIs, browsers, or retrieval where authorization, retries, ambiguous outcomes, or runtime identity are difficult to prove. Email axiom.co@proton.me with the workflow, operating environment, tool/effect boundary, known failure modes, and current validation. AXIOM will confirm fit before any work begins.
Request sprint fit reviewLimits
Engineering evidence, not certification.
This engagement is not independent assurance, not certification, not a legal or compliance opinion, and does not imply a production deployment. It does not establish exactly-once external effects. Production deployment, provider migration, paid infrastructure, credentials administration, ongoing monitoring, and materially expanded scope are separate work.