Senior AI Agent Engineer

Python, FastAPI and Microsoft Agent Framework Orchestration | FreelanceElchai Group | Bounded contract opportunity | Remote working by agreementElchai is developing New Pulse as the authenticated human interface to a governed business-assistant architecture. Selection and any paid work remain subject to approval. This is not an unpaid internship or an invitation to start work before a contract.One approval-paused, recoverable workflow packageIntegrate one deterministic workflow with the approved existing orchestration runtime and control-plane contracts. Demonstrate a real pause, verified resume, isolated checkpoint restoration, bounded retries and duplicate prevention against an offline recording adapter.This is one workflow integration into an approved runtime. It excludes building the entire control plane, identity system, evidence platform or a second orchestrator.We agree the exact scope, baseline or input register, interfaces, tests, reviewer, fees and availability before work. For development, the assignment uses an isolated sandbox and synthetic data; internal owners retain architecture, security, independent QA and release authority. For research/commercial work, only authorized public-source research and synthetic exercises are included unless separately stated.Milestones and smaller deliverablesM0: Freeze the one-workflow integration (client readiness gate; no free candidate deliverable).M0.1: prove runtime/reuse. M0.2: freeze contracts and fixtures. M0.3: approve files, tests, owners and start.Required evidence: Engine SHA/runtime; verified dependencies; workflow/envelope/decision hashes; test plan and capacity.Pass condition: The client confirms the agreed package, necessary dependencies, named reviewer and permission to start. A quote or interview does not require unpaid implementation.M1: Implement four small workflow pieces (16-24 focused execution hours).M1.1: compatibility and adapter seam, 4-6 h. M1.2: pause/verified resume, 4-6 h. M1.3: durable checkpoint binding, 4-6 h. M1.4: bounded retry/idempotent effect, 4-6 h.Required evidence: Four checkpoint handovers, actual runtime traces, synthetic state fixtures and recording-adapter log.Pass condition: Real runtime executes the agreed path without reaching external providers or bypassing approval.M2: Verify authority, recovery and duplicate prevention (12-20 focused execution hours).M2.1: authority/version negatives, 4-6 h. M2.2: kill/restart/fault cases, 4-8 h. M2.3: duplicate/tool/evidence checks, 4-6 h.Required evidence: AE-01 to AE-08; raw failure/recovery traces; exact versions; concurrency and duplicate counts.Pass condition: All frozen criteria pass; zero unauthorized or duplicate effects; no unresolved control/security defect.M3: Independently replay the workflow (4-8 focused execution hours).M3.1: reviewer runs the state paths. M3.2: corrections/retest, 4-8 h. M3.3: accept exact workflow package.Required evidence: Demo from clean checkpoint; reviewer replay; rollback; accepted limitations and takeover smoke.Pass condition: Exact package accepted as sandbox-only; no claim of full T11, live mail or engine productization.M4: Transfer runtime knowledge and actuals (2-4 focused documentation/reporting hours).M4.1: workflow/adapter/runbook handover. M4.2: reconcile effort, defects and next workflow proposal.Required evidence: State transition table, checkpoint formats, dependencies, actuals, remaining core gaps.Pass condition: Another approved developer can replay and diagnose the flow; no automatic next workflow award.The hour ranges are initial planning guides, not deadlines or accepted prices. Quote your own effort and fee by milestone, including documentation, correction assumptions and availability. Internal approval and review time is separate; waiting for those decisions is not concealed as development work.Verification focusDemonstrate that nothing executes before a verified decision; a changed action cannot use an earlier approval; a restart restores the right branch state; and a duplicate/retry does not repeat an effect. The proposed scenario set includes 50 approval/rejection/change/forgery paths and separate restart/duplicate tests. Missing core runtime dependencies block the package rather than expand its scope.Handover at every checkpointDeliver editable files, the exact current version, setup/test commands or research method, interface/source definitions, known defects, accepted versus unfinished work and the next task. Another authorized contributor must be able to reproduce the checkpoint without relying on your private account or undocumented knowledge. Final handover includes a clean-start reproduction and measured hours/corrections. We do not request confidential material from former clients.Experience evidenceRelevant skills: Production Python/FastAPI; Microsoft Agent Framework or deeply comparable workflow experience; durable state; verified human decision integration; idempotency and bounded retries.Provide two relevant shipped-work or redacted project examples and your precise contribution; answer this question: Explain a restart or uncertain-outcome failure you solved. How did you keep approval valid, preserve state and avoid repeating the side effect?Include your proposed milestone fees, currency, availability, working-time overlap, assumptions, exclusions and any subcontracting. We do not request free test work.