Recommendation Continue the Autonomous Project Orchestrator, but do not implement the entire execution stack yourself. None of the established FOSS projects is a complete replacement for the design described in the supplied README.md. The closest projects solve valuable subsets, particularly durable execution, coding-agent loops, workflow scheduling, sandboxing and observability. The strongest direction is: 1. Retain your deterministic control plane, work graph, evidence model, budget accounting and integration semantics. 2. Use Temporal as the leading candidate for durable activity execution and crash recovery, after a focused compatibility prototype. 3. Support OpenHands, SWE-agent and Aider as interchangeable coding executors, rather than building a coding agent from scratch. 4. Use Kubernetes Jobs or Argo Workflows for isolated, resource-limited execution where Kubernetes is available. 5. Adopt OpenTelemetry for observability and OPA for policy evaluation. 6. Keep PostgreSQL as the authoritative event, work-graph and evidence store, unless the implementation proves that Temporal can safely own part of that state. This is a recommendation to continue with a narrower scope, not to replace the project or to build every component. Transparency: I was able to assess the README content returned in the conversation, but I do not have a functioning live web-search facility in this response. The licence and capability information below reflects the projects’ established public repositories up to my available knowledge and must be rechecked against each project’s current LICENSE, release notes and contributor activity before procurement or integration. Why No Existing Project Fully Replaces It Your scope is materially broader than “run several agents”: - Recursive problem decomposition - Parallel execution - Multi-model reasoning - Deterministic tools - Integration and conflict ownership - Validation backed by evidence - Hierarchical budgets and accounting - Replanning - Explicit state machines - Event catalogue and replay behaviour - Crash recovery - Executor trust boundaries - Conformance testing Most agent frameworks concentrate on: - Message passing - Prompt and tool orchestration - Agent hand-offs - LLM provider integration - Memory - Human approval points They typically do not provide a sufficiently precise model for repository integration, acceptance evidence, conflict classification, hierarchical budget conservation, deterministic reduction, or project-level run outcomes. Conversely, workflow engines provide durable state and retries but do not understand software-development semantics. Coding agents can edit and test repositories, but they do not provide the governing control plane. Your differentiator is therefore not “multiple agents”. It is the deterministic governance of untrusted, probabilistic executors working on a shared project. Candidate Assessment Project Licence to verify Strong fit Important mismatch Recommended role Temporal MIT Durable workflows, retries, timers, activity execution, recovery Workflow replay constraints; not a project/work-graph model; no repository integration semantics Durable command and activity substrate Argo Workflows Apache-2.0 Kubernetes-native DAGs, artefacts, retries, resource controls Static/container-oriented workflows; weaker dynamic governance model Sandboxed execution backend Flyte Apache-2.0 Typed workflows, caching, artefacts, Kubernetes execution Data/ML orientation; heavyweight for coding-agent control Alternative execution backend Apache Airflow Apache-2.0 Mature scheduling, operators, observability Primarily scheduled batch DAGs; dynamic autonomous replanning is awkward Not recommended as core Prefect Apache-2.0 core, verify current boundaries Python-native workflows and operational UI Product boundary and server/cloud feature allocation need checking; not project governance Possible prototype backend LangGraph MIT Stateful agent graphs, checkpoints, interrupts, human approval LLM-centric; weaker formal integration, accounting and conformance semantics Planner/executor implementation library Microsoft AutoGen MIT, verify current repository/version Multi-agent communication and tool use Message-centric; does not replace durable governance Optional executor adapter AG2 Apache-2.0, verify AutoGen-derived agent orchestration Similar message-centric mismatch Optional executor adapter Semantic Kernel MIT Model/provider abstractions, plugins, planning primitives SDK rather than durable orchestrator Provider and tool adapter library CrewAI MIT, verify current editions Rapid role/task-based agent workflows Higher-level abstractions; limited deterministic project governance Experiments only OpenHands MIT Autonomous software-work loops, repository editing, runtime isolation It is principally an executor, not the global control plane Primary coding-executor adapter SWE-agent MIT Focused issue-resolution loop and benchmark heritage Narrow task loop; lacks portfolio/work-graph governance Coding-executor adapter and baseline Aider Apache-2.0 Mature repository editing and model support Interactive coding agent, not orchestration control plane Lightweight coding-executor adapter Open Policy Agent Apache-2.0 Deterministic policy decisions and auditable policy bundles Does not manage workflows Policy-evaluation component OpenTelemetry Apache-2.0 Vendor-neutral traces, metrics and logs Not an event ledger or evidence model Mandatory observability layer Kubernetes Apache-2.0 Isolation, quotas, jobs, scheduling and ecosystem Operationally expensive; not sufficient as a sandbox against hostile workloads by itself Execution substrate Podman Apache-2.0 Rootless containers and local execution Host/kernel isolation limitations Local executor backend gVisor Apache-2.0 Stronger container isolation Compatibility and performance trade-offs Untrusted executor sandbox Firecracker Apache-2.0 MicroVM isolation More operational integration work High-trust-boundary deployments NATS Server/JetStream Apache-2.0 Messaging, persistence and work distribution Not an authoritative relational project model Optional event transport PostgreSQL PostgreSQL Licence Transactions, event ledger, graph state, evidence metadata Requires application-level orchestration logic Recommended source of truth Best Replacement Candidates Temporal Temporal is the most credible candidate for replacing a meaningful amount of infrastructure: - Durable activity invocation - Retry and timeout handling - Timers - Cancellation propagation - Worker failure recovery - Long-running execution - Idempotency-oriented activity design - Operational visibility However, there is a significant architectural issue: Temporal replays workflow code to reconstruct state and therefore restricts nondeterministic operations inside workflow definitions. Your design already has a deterministic reducer and an event catalogue. Combining two replay authorities carelessly would produce: - Duplicate state models - Ambiguous ownership of retries and cancellation - Difficult schema/version migration - Two event histories with different semantics - Potential disagreement after recovery - Coupling of normative control-plane transitions to Temporal implementation details Recommended Temporal boundary: the orchestrator remains authoritative. It emits commands such as “execute node attempt”, “run validation” or “prepare integration”. Temporal durably executes those commands and reports immutable results. The orchestrator reducer consumes the results and decides the next normative state. Do not initially encode the entire orchestrator state machine as one Temporal workflow. OpenHands OpenHands appears to be the strongest candidate for avoiding implementation of a complete coding-agent loop. It can potentially provide: - Repository interaction - Shell/tool execution - Model-driven editing - Runtime isolation - Iterative test-and-fix behaviour - Existing model/provider support Treat it as an untrusted executor behind your executor protocol. It should not directly determine that work is accepted, integrated, complete or sufficiently evidenced. LangGraph LangGraph is useful where a planner or executor requires an internal graph with checkpoints and human intervention. It is not an adequate authority for your project state machines. A sensible boundary is: Deterministic orchestrator -> node attempt command -> LangGraph-based planner or executor -> proposed actions and artefacts <- structured result -> independent validation and reducer transition Avoid allowing LangGraph checkpoints to become the authoritative record for project state. Argo Workflows Argo is attractive if the deployment already assumes Kubernetes. It can run isolated per-attempt containers with: - CPU and memory limits - Deadlines - Service accounts - Network policies - Volume and artefact controls - Retry policies Argo is less suitable as the primary control plane because autonomous decomposition and replanning continuously mutate a semantic work graph. Encoding that behaviour into generated Argo DAGs would make recovery and audit semantics harder to reason about. What to Build Keep these components project-specific: - Deterministic reducer - Normative state machines - Work-graph contract - Event catalogue and event schemas - Node acceptance contract - Evidence binding and freshness - Hierarchical budget allocation and accounting - Replanning policy - Executor-selection policy - Repository integration ownership - Conflict taxonomy and resolution state - Run outcome calculation - Conformance harness - Recovery reconciliation - Human intervention and override audit trail These are the design’s differentiators and are poorly covered by existing projects. What Not to Build Adopt existing implementations for: - Container and microVM execution - Kubernetes scheduling - Generic durable retries and timers - LLM provider SDKs - OpenTelemetry instrumentation - Policy-language parsing and evaluation - Git plumbing - Secret storage - Identity and access management - Message transport - General dashboards and log collection - Coding-agent edit loops Writing private replacements for these components would increase security risk and delay validation of the distinctive architecture. Proposed Reference Architecture API / CLI / operator interface | v Deterministic control-plane service - command validation - reducer - state machines - budgets - policy outcomes - recovery reconciliation | +----------------------+ | | v v PostgreSQL event ledger OPA policy engine - events - admission policy - work graph - executor selection constraints - attempts - budget and trust rules - evidence - integration state | v Durable dispatch layer - Temporal initially - or PostgreSQL transactional outbox plus workers | v Executor adapters - OpenHands - SWE-agent - Aider - deterministic tool runner - human executor | v Sandbox backends - rootless Podman for development - Kubernetes Jobs plus gVisor for production - Firecracker where stronger separation is needed | v Validation and integration workers - deterministic checks - independent model evaluation - Git integration owner - evidence production Cross-cutting: OpenTelemetry traces, metrics and correlated event IDs Critical Design Principle The execution system may report facts, but it must not decide normative state. For example: { "attempt_id": "attempt-123", "command_id": "command-456", "executor": "openhands", "exit_status": 0, "artefacts": ["patch.diff", "test-results.xml"], "started_at": "...", "finished_at": "..." } That result must not directly mean ACCEPTED, COMPLETED or INTEGRATED. The reducer should derive the transition only after checking: - Command identity and idempotency - Attempt freshness - Budget ledger - Required evidence - Validator independence - Acceptance criteria - Integration ownership - Current node state - Cancellation or supersession - Policy outcome This separation is where your orchestrator adds value beyond the alternatives. Decision Matrix Using a five-point qualitative scale: Criterion Continue custom Temporal alone LangGraph OpenHands Argo Exact documented semantics 5 2 2 1 1 Durable execution 2 initially 5 3 2 4 Coding autonomy 2 initially 1 3 5 1 Repository integration governance 5 potential 1 1 3 1 Budget and evidence semantics 5 potential 1 2 1 1 Crash recovery 3 potential 5 3 2 4 Implementation effort 1 4 4 4 3 Avoidance of lock-in 4 3 3 3 4 Conformance to formal specification 5 potential 2 2 1 1 No candidate dominates the custom design on its differentiating criteria. The custom design is weakest exactly where mature infrastructure should be adopted: durable dispatch, sandboxing and coding-agent implementation. Recommended MVP Boundary The first implementation should demonstrate the core hypothesis rather than implement every document. Include: - One process-level control-plane service - PostgreSQL event ledger - Deterministic reducer - A small subset of work-node and attempt states - Transactional outbox - One deterministic executor - One coding executor, preferably OpenHands or Aider - Rootless container sandbox - Git worktree isolation - Deterministic validation - Evidence attachment - Simple hierarchical token and wall-clock budgets - One integration owner - Crash and duplicate-delivery tests - OpenTelemetry instrumentation Defer: - Multiple workflow backends - Sophisticated model routing - Automatic semantic merge-conflict resolution - Distributed consensus - Dynamic marketplace of executors - Custom scheduler - Custom sandbox technology - Broad UI - Cross-project optimisation Required Proof Before Committing to Temporal Implement a narrow spike containing: 1. A node attempt dispatched as a Temporal activity. 2. Worker termination midway through execution. 3. Duplicate completion delivery. 4. Cancellation racing with successful completion. 5. Control-plane restart. 6. A superseded attempt returning after a newer attempt. 7. Event-schema and reducer-version upgrade. 8. Reconciliation between Temporal history and the orchestrator ledger. Success criteria: - PostgreSQL and the reducer remain authoritative. - Every externally visible result is idempotent. - A late result cannot revive a cancelled or superseded attempt. - Temporal retry state does not consume the budget twice. - Recovery requires no manual history editing. - An execution backend can later be replaced without changing normative event meanings. If this boundary proves awkward, use a simpler PostgreSQL transactional outbox plus leased workers for the MVP. Your state model is already explicit enough that this may be simpler than operating two durable state systems. Licensing Gate Before adopting any candidate, record: - Exact repository URL - Exact release or commit - SPDX licence identifier - Copyright notice obligations - Dependency licence scan - Whether server and client use different licences - Whether enterprise plugins are required - Whether trademarks restrict redistribution - Whether hosted functionality is absent from the FOSS release - CLA requirements for upstream contributions - Source-available or Business Source Licence components Exclude any required component that is merely “source available”. FOSS acceptance should follow an OSI-approved licence, not just public source code. Final Decision Proceed with the project, but redefine it explicitly as a deterministic governance and integration control plane, not another agent framework or workflow engine. The preferred stack is: - Custom: reducer, event semantics, work graph, budgets, validation evidence, integration, conflicts and conformance - Temporal: candidate durable execution layer - OpenHands plus Aider or SWE-agent: coding executor implementations - PostgreSQL: authoritative ledger and state - OPA: policy evaluation - Podman/Kubernetes plus gVisor: execution isolation - OpenTelemetry: observability - Git worktrees: repository isolation A wholesale adoption of LangGraph, AutoGen, CrewAI, OpenHands, Argo or Temporal would discard the project’s central value without satisfying its normative semantics. A wholly custom implementation would waste effort and create avoidable reliability and security risks. The hybrid approach preserves the genuinely novel control-plane work while sourcing mature commodity subsystems from established FOSS projects. The model hit its output limit, so this response may be incomplete.