Durable Agentic Harness — crash-safe autonomous AI agents with human-in-the-loop approval

[darshitvvora]

Project link

https://github.com/darshitvvora/durable-agentic-harness

Language

Python

Short description (max 256 chars)

An autonomous stock-trading agent (OpenAI Agents SDK) that reframes Temporal as the Durable OS for agentic AI: workers = scheduling, event history = autosave, signals = human-in-the-loop. Kill the worker mid-trade and the agent replays from the exact line.

Long Description

Temporal: The Durable OS for Agentic AI

Agents are easy to demo, hard to operate. Every production agent eventually hits the
same wall — LLM calls flake, workers crash mid-tool-call, human approvals stall for
hours, parallel work loses children on restart. The usual answer ("just add retries +
Redis + a state machine") is the long road to badly reinventing Temporal.

This demo reframes Temporal not as a workflow engine but as OS primitives for agent
loops, underneath an autonomous OpenAI-Agents-SDK stock-trading agent:

OS primitive	Temporal equivalent	What it gives agents
Process scheduling	Workers + task queues	LLM/tool work dispatched durably
Autosave / journaling	Event history	Replay from the exact event after a crash
IPC / interrupts	Signals, Updates & queries	Human-in-the-loop, mid-flight steering
Memory / state	Workflow state	Survives restarts — no Redis, no S3 checkpoints
Drivers	Activities	Side-effects, retried & idempotent by default
Long-lived sleep	workflow.sleep()	Pause days/weeks at zero CPU cost

Who this is for / use cases:

The trading agent is the vehicle — the real subject is the pattern for any
long-running, autonomous, or human-supervised agent. Reach for this when:

• Crash-safe agent loops — agents that run for minutes to weeks and must survive
  worker restarts, deploys, and infra failures without losing in-flight state or
  re-doing side-effects (orders placed, emails sent, payments made).
• Human-in-the-loop approvals — workflows that pause indefinitely at zero CPU cost
  waiting on a human to approve/reject a high-stakes action (a large trade, a refund,
  a production change), then resume exactly where they left off.
• Parallel fan-out / fan-in — exploring N strategies, prompts, or candidates
  concurrently in isolated sandboxes and selecting a winner, with automatic cleanup of
  children if the parent restarts.
• Auditable AI decisions — every LLM call, tool call, and signal is a queryable
  event in history, giving you a replayable audit trail for compliance and debugging
  ("why did the agent do X at tick 14?").
• LLM/tool reliability — wrapping flaky model and API calls as activities so they
  retry idempotently by default, instead of hand-rolling retry + backoff logic.

If you're building agentic systems and finding yourself bolting on Redis, Celery, a
retry library, and a state machine for approvals, this demo shows what those concerns
look like when Temporal owns them instead.

What the agent does:

1. Discovers a strategy by fanning out N parallel sandboxed backtests in airgapped
   Docker containers (child workflows).
2. Lives through a tick loop: market + news context, LLM trade-intent via the OpenAI
   Agents SDK (activity_as_tool), a deterministic risk guardrail, and human-in-the-loop
   approval for large trades.
3. Survives chaos — kill the worker mid-trade and Temporal replays the decision
   history to resume from the exact line.

Stack: FastAPI (sole Temporal client) + SSE, React 18 / Vite / Tailwind UI,
temporalio[openai-agents] workflows wired via OpenAIAgentsPlugin, Docker sandboxes
for backtests, and Mockoon for offline/deterministic market·news·broker data.

What it strips out — that you'd otherwise write yourself: no Celery, no Redis-backed
queue, no hand-rolled retry policy, no "save progress to S3" code, no state machine for
approvals, no orphan-child cleanup. Temporal owns all of it; what's left on top is just
the agent logic.

What Linux did for processes, Temporal does for agent loops.

Screenshots and architecture/sequence diagrams are in the repo README.

Author(s)

Darshit Vora - Staff Solutions Architect @ Temporal
LinkedIn - linkedin.com/in/darshitvvora

[github-actions]

Hi, I'm ZiggyBot! 🤖 Here's my pre-evaluation of this submission:

---

Criterion	Assessment
Useful to Temporal users	✅ Yes — Demonstrates a compelling pattern for building crash-safe AI agents with human-in-the-loop approval, directly addressing real production challenges in agentic AI systems.
Benefits clearly explained	✅ Yes — The "Temporal as Durable OS" framing with the OS primitives table is exceptionally clear, and the README thoroughly explains what problems this solves and what boilerplate it eliminates.
OSI Approved License	⚠️ Unclear — No license file was detected by GitHub; the project needs an explicit LICENSE file to meet open-source requirements.
README with working instructions	✅ Yes — The README includes architecture diagrams, a detailed sequence diagram, and appears to have comprehensive setup instructions (though truncated in the review, the structure suggests completeness).

Overall: Needs review ⚠️

Notes:

• This is a high-quality demonstration project with excellent documentation and a thoughtful conceptual framing that could genuinely help developers understand how Temporal applies to agentic AI workloads.
• The integration with the OpenAI Agents SDK via temporalio[openai-agents] and the OpenAIAgentsPlugin showcases a modern, relevant use case.
• The fan-out pattern for parallel backtests in Docker sandboxes and the signal-based human-in-the-loop approval flow are well-implemented Temporal patterns.
• The only blocking issue is the missing license — this needs to be resolved before acceptance into the Code Exchange.
• The project has only 1 star, but quality matters more than popularity for a curated showcase.

Suggested questions for submitter:

• Could you add an OSI-approved license file (MIT recommended) to the repository?
• Is there a deployed demo or video walkthrough that shows the chaos resilience (kill-worker-mid-trade replay) in action?

Teaching moment:

The BacktestSandboxWorkflow demonstrates a clean pattern for wrapping potentially flaky, long-running work (Docker sandbox execution) as a Temporal activity with appropriate timeouts and retry policies. Notice how the workflow gracefully degrades by returning a Scorecard with an error field rather than failing entirely — this allows the parent workflow's fan-out to continue even if some backtests fail. The combination of start_to_close_timeout, heartbeat_timeout, and retry_policy shows production-ready activity configuration for unpredictable external processes.
View in source ↗

---

---

ZiggyBot is an AI pre-screener based on Temporal's community mascot Ziggy. Final decisions are made by the community team.

Reviewer: Check the box below to generate a Contentful-ready long description for this submission.

• Generate long description if needed

[github-actions]

Hi, I'm ZiggyBot! 🤖 Here's my pre-evaluation of this submission:

---

Criterion	Assessment
Useful to Temporal users	✅ Yes — Demonstrates a compelling "Temporal as Durable OS for AI agents" pattern with crash recovery, human-in-the-loop signals, and fan-out/fan-in child workflows that directly addresses real production challenges for agentic AI systems.
Benefits clearly explained	✅ Yes — The README provides an excellent OS-primitive mapping table, detailed architecture diagrams, sequence flows, and clearly articulates what Temporal replaces (Celery, Redis, hand-rolled retry, state machines).
OSI Approved License	⚠️ Unclear — No license file was detected by GitHub; the project needs an explicit OSI-approved license (MIT recommended) to meet Code Exchange requirements.
README with working instructions	✅ Yes — The README includes architecture diagrams, a comprehensive sequence diagram, and based on the truncated content appears to have setup instructions, though the full quickstart section was cut off.

Overall: Needs review ⚠️

Notes:

• This is a high-quality, well-documented submission that showcases advanced Temporal patterns including the new temporalio[openai-agents] integration with OpenAIAgentsPlugin.
• The conceptual framing of "Temporal as a Durable OS" is pedagogically valuable and could resonate well with developers coming from traditional infrastructure backgrounds.
• The project demonstrates multiple Temporal features in concert: child workflows for parallel backtests, signals for human approval, activities for side effects, and crash recovery via event history replay.
• The only blocking issue is the missing license file — this must be addressed before acceptance.
• The 1-star count is expected for a new submission and shouldn't be a concern.

Suggested questions for submitter:

• Could you add an OSI-approved license file (MIT recommended) to the repository root?
• The README was truncated — can you confirm that complete setup instructions (including Mockoon configuration, Docker requirements, and environment variables) are present?
• Have you tested the crash-recovery scenario end-to-end, and does replay work correctly with the OpenAI Agents SDK plugin across worker restarts?

Teaching moment:

The BacktestSandboxWorkflow elegantly demonstrates how to wrap potentially long-running, side-effect-producing work (Docker sandbox execution) as a Temporal activity with appropriate timeouts and retry policies. By returning a Scorecard with an error field on failure rather than raising, it ensures the parent workflow can gracefully handle partial failures during fan-out without losing completed results. This pattern of "fail gracefully, don't crash the workflow" is crucial when orchestrating parallel work where some children may fail.
View in source ↗

---

---

ZiggyBot is an AI pre-screener based on Temporal's community mascot Ziggy. Final decisions are made by the community team.

Reviewer: Check the box below to generate a Contentful-ready long description for this submission.

• Generate long description if needed

[github-actions]

Hi, I'm ZiggyBot! 🤖 Here's my pre-evaluation of this submission:

---

Criterion	Assessment
Useful to Temporal users	✅ Yes — Demonstrates a compelling "Temporal as durable OS for AI agents" pattern with crash recovery, human-in-the-loop approvals, and parallel child workflows that addresses real production challenges for agentic AI systems.
Benefits clearly explained	✅ Yes — The README provides an excellent conceptual framework mapping OS primitives to Temporal equivalents, clear architecture diagrams, and detailed sequence flows explaining what Temporal handles versus what you'd otherwise build yourself.
OSI Approved License	⚠️ Unclear — No license file was detected by GitHub; the project needs to add an explicit OSI-approved license (MIT recommended) before acceptance.
README with working instructions	⚠️ Unclear — The README was truncated, so complete setup instructions couldn't be verified; the visible portions show strong documentation structure but we can't confirm if Docker setup, environment variables, and run commands are fully documented.

Overall: Needs review ⚠️

Notes:

• This is a high-quality, thoughtfully designed demonstration of using Temporal with the OpenAI Agents SDK. The conceptual framing ("Temporal as a Durable OS") is genuinely valuable for helping developers understand Temporal's benefits.
• The temporalio[openai-agents] integration with OpenAIAgentsPlugin showcases the newest Python SDK capabilities for AI agent orchestration.
• The architecture includes sophisticated patterns: fan-out child workflows for parallel backtests, signals for human-in-the-loop approval, and Docker sandboxing for isolated execution.
• The single star count suggests this is a new project, which is fine for the Code Exchange.
• The missing license is a blocker that must be resolved before acceptance.

Suggested questions for submitter:

• Could you add an OSI-approved license file (MIT is recommended) to the repository?
• Is the full README available with complete setup instructions, including environment configuration, Docker requirements, and how to run Mockoon for the mocked services?
• Are there any additional dependencies or prerequisites (OpenAI API key, specific Docker version) that users should be aware of?

Teaching moment:

The BacktestSandboxWorkflow demonstrates a clean child workflow pattern where each backtest runs as an isolated workflow with its own retry policy and timeout. Notice how the workflow gracefully degrades on failure by returning a Scorecard with an error field rather than propagating the exception—this allows the parent workflow to collect partial results from successful backtests even if some fail. This pattern of "fail gracefully, report errors in results" is excellent for fan-out scenarios where you want to maximize successful completions.
View in source ↗

---

---

ZiggyBot is an AI pre-screener based on Temporal's community mascot Ziggy. Final decisions are made by the community team.

Reviewer: Check the box below to generate a Contentful-ready long description for this submission.

• Generate long description if needed

[clementineCU]

This is a strong durable-agent pattern. One small extension that would make it useful for market-agent evaluation: publish a one-market replay bundle alongside the demo.

Minimal shape: source snapshot + timestamp/block, exact command or prompt, observed/inferred/not-checked labels, trade or no-trade decision, risk boundary, and flip condition.

We are using that artifact shape for Bazaar Signal Agent submissions here: https://github.com/awrsla/agent-bazaar-challenge

[github-actions]

Hi, I'm ZiggyBot! 🤖 Here's my pre-evaluation of this submission:

---

Criterion	Assessment
Useful to Temporal users	✅ Yes — Demonstrates a compelling pattern for building crash-safe AI agents with human-in-the-loop approval using Temporal's durability guarantees, directly applicable to production agentic systems.
Benefits clearly explained	✅ Yes — The "Temporal as Durable OS for Agentic AI" framing with the OS primitives mapping table is exceptionally clear, and the submission articulates specific use cases (crash-safe loops, human approvals, parallel fan-out, audit trails).
OSI Approved License	⚠️ Unclear — GitHub reports "Other (SPDX: NOASSERTION)" which suggests a license file may exist but isn't recognized; needs verification that it's an OSI-approved license.
README with working instructions	⚠️ Unclear — The README excerpt is truncated and while it shows excellent architecture documentation, the actual setup/installation instructions weren't visible in the provided content.

Overall: Needs review ⚠️

Notes:

• This is a high-quality, well-documented submission with excellent pedagogical value for the Temporal community. The "OS primitives" framing is genuinely novel and helps developers understand Temporal's value proposition for AI workloads.
• The integration with OpenAI Agents SDK via temporalio[openai-agents] and OpenAIAgentsPlugin demonstrates cutting-edge Temporal patterns.
• Architecture diagrams (Mermaid) and sequence diagrams are thorough and professional.
• The backtest workflow code sample shows clean, idiomatic Temporal Python usage with appropriate retry policies and heartbeats.
• The license issue needs clarification before acceptance — please confirm the license type in the repository.

Suggested questions for submitter:

• Could you confirm which OSI-approved license this project uses? GitHub isn't recognizing the license file.
• Are the setup instructions (Docker, Mockoon, environment variables, etc.) complete in the full README? The truncated version didn't show the installation steps.
• Does the demo work with Temporal Cloud as well as local temporal server start-dev?

Teaching moment:

The child workflow pattern for parallel backtests is a great example of combining Temporal's durability with fault-tolerant fan-out. The BacktestSandboxWorkflow wraps a potentially long-running Docker sandbox execution as an activity with a 180-second timeout, heartbeat support, and a 2-attempt retry policy — returning a graceful error Scorecard instead of failing the entire workflow on exception. This pattern shows how to safely orchestrate N parallel exploratory tasks while maintaining parent workflow resilience.
View in source ↗

---

---

ZiggyBot is an AI pre-screener based on Temporal's community mascot Ziggy. Final decisions are made by the community team.

Reviewer: Check the box below to generate a Contentful-ready long description for this submission.

• Generate long description if needed