Four open-source runtimes prove the full autonomous loop — one, temporal-community's refund agent, closes it against a real payment owner
Asked (summary):
Find open-source agent runtimes, workflow engines, and agent-evaluation systems that prove an autonomous loop end to end — a real worker or scheduler wakes, a new decision is made after a checkpoint, the decision persists across restart, an external/tool side effect is verified, and retained-vs-removed learning is compared under equal future inputs — with primary, executable evidence per implementation: exact file paths, checkpoint/fork method, proof of a genuinely new forward decision, worker trigger, idempotency handling, causal trace fields, independent verification, and the pattern most relevant to fixing OpenPresence. Exclude leaderboards, prompt-only tests, and claims without runnable evidence.
A strict composite screen — runnable evidence only — found four open-source implementations covering the requested proof chain: two Temporal-based agent runtimes, a Temporal community refund-agent demo, and an Inngest self-learning agent. Research was conducted 23 September 2026; the evidence is not uniform in strength, and the matrix below marks where verification or ablation is weaker.
Proof-chain matrix — six stages per implementation
strong executable, primary evidence
caveat evidence present but weaker — see the row's detail
Click a row to expand file paths, commands, causal identifiers, and the OpenPresence fix pattern
project
worker wake
checkpoint / restart
new decision
retry-safe effect
independent verification
equal-input ablation
Ranked by completeness of the external-side-effect proof. temporal-community/agent-memory-and-state is the most complete: its verifier is the effect owner itself (Stripe), and its retained-vs-removed comparison runs the identical refund request twice.
What the badges cannot say
temporal-community's replacement Worker resumes at the saved "issue refund" step and actually executes the refund against Stripe — one Workflow identity maps to one idempotency key, so a retry after a simulated timeout resolves two calls to one refund.
github.comInngest's agent re-reads its prompt registry on every loop start and makes a fresh weighted-random version selection; promotion demands a ≥1.0-point composite advantage at ≥80% traffic, with baseline v1 never deleted — a permanent control.
github.comtemporal-agent-sdk's stateful MCP test is a live ablation: one session's thoughtHistoryLength climbs across calls, whereas a reconnect per call would reset it to 1 — retained versus discarded memory under equal inputs.
github.comagent-sdk-go's scenario 5c is the most explicit kill test — pkill -SIGKILL mid-stream, then reconnect from saved offset 10 through completion — but its verifier is operational inspection (Temporal UI, state file), not a separate effect owner.
pkg.go.dev
Synthesis for OpenPresence
- Build a crash-injection acceptance test around a durable scheduler. Kill the worker mid-loop and require a replacement worker to reconstruct state and continue — Temporal event-history replay or Inngest step checkpoints, proven by a scripted kill/restart scenario.
- Record the parent checkpoint ID and a fresh post-resume decision/event ID. The proof of a new forward decision is a new identifier created after resume, causally linked to the checkpoint it resumed from.
- Map workflow/run identity to an effect idempotency key. One loop identity becomes one external-effect key, so a retry after an uncertain call asks the effect owner instead of guessing.
- Verify through the external effect owner. Confirm the side effect where it lives — as the refund demo queries Stripe directly — not through the agent's own logs.
- Replay identical future inputs against retained-learning and removed-learning forks, with a permanent control. Attribute outcomes per version and never delete the baseline, only deprioritize it.
All four rows
| Project | File path | Checkpoint method | Worker trigger | Independent verifier | Source |
|---|
Strict evidence screen of open-source agent runtimes, workflow engines, and agent-evaluation systems, conducted 2026-09-23: 4 implementations, one row each, kept only when runnable evidence covers the full loop (worker wake, checkpoint/restart, new post-restart decision, retry-safe side effect, independent verification, equal-input retained-vs-removed comparison). Badges grade evidence strength per stage; full verbatim evidence text lives in each row's expandable detail. Not an exhaustive census — documentation-only durability claims, leaderboards, and prompt-only tests were rejected. Table trims retry, trace, decision-proof, ablation, and command fields for width; expand a matrix row for those.