Agentic workflow platform

Author. Run.
Diagnose. Improve.

The full lifecycle of agentic workflows — from versioned DAG definitions to per-node diagnosis, with evals that measure improvement by the numbers, not by feel.

Inspect each node: exact prompt, context, model, token usage
Compare eval scores before and after prompt iterations
Provider-agnostic — swap models without changing the DAG
Sign in to workspace
SDLC Theme · story/ARC-427run_9f2k4a · 4m 32s elapsed
Active
context:ARC-427.work_item
model:sonnet-4-6
provider:anthropic
rag:codebase · 3 docs
tokens:8.2k in · 2.1k out
theme:sdlc · v2.1.0
implement· tool calls
read_filepackages/core/ports/runs.py1.2s
read_fileapps/api/features/runs/models.py0.8s
read_filedocs/decisions/ADR-0018.md0.4s
write_fileapps/api/features/runs/routes.py
context surface
work_itemARC-427 · Add run action endpoints
rag · 3 docsADR-0018 · runs/ports.py · runs/models.py
upstreamplan.output · scope.output
promptsdlc/implement.jinja2 · v4
18:42:03[scope]Work item resolved · 9 tasks scoped
18:44:17[plan]Architecture plan complete · ADR-0018 applied
18:46:31[impl]RAG resolved · 3 context docs attached
18:47:02[impl]3 files read · writing routes.py
18:47:09[impl]POST /runs/{id}/actions · 47 lines
3
Active Runs
94%
Pass Rate
12
Stories
8.2k
Avg Tokens
Recent Runslast 24h · 8 total
▶run_9f2k4aARC-427 · Implement run actions4m 32s● running
▶run_8k1p3eARC-425 · WorkGraph coordinator2m 11s● running
✓run_7m3k1bARC-421 · Implement runs routes3m 18s9m ago
✓run_5n8j2cARC-419 · Fix auth middleware5m 44s22m ago
✓run_3p4h7dARC-415 · Add eval integration4m 08s1h ago
✓run_2q9r6fARC-411 · Tracing adapter6m 22s2h ago
WorkGraphSprint 4
Sprint 4 · Foundation8 / 12 complete
2 in-progress2 queued0 blocked8 done
Quality Gateslast 12 runs · devloop stage
vendor-boundary
12/12✓
import-linter
12/12✓
banned-patterns
12/12✓
eval-pass-rate
10/12↗
pyright-strict
11/12↗
Provider Mixlast 30d
anthropic · sonnet-4-6
72%
anthropic · haiku-4-5
20%
openai · gpt-4o
8%
Eval score progression · sdlc/implement.jinja2
What changed
v1
58%✗missing context scope — no work_item in prompt surface
v2
71%✗added RAG context — codebase + ADR-0018 attached
v3
84%✗tightened output schema — strict JSON mode enforced
v4
94%✓injected upstream task outputs — plan.output threaded in
Eval dimensions · v4 score breakdown
code-correctness
96%
output-schema
97%
test-coverage
92%
arch-adherence
88%
token-efficiency
91%
+36ppeval score58% → 94%
−18%token cost4.1k → 3.4k avg
+2/12pass ratewas 7/12 → 10/12
● v4 promoted to production · eval: implement-quality · baseline: v1 · 2h ago
The improvement loop

Every step captured.
Every failure diagnosable.

The closed loop that separates a governed platform from a pile of scripts.

01

Author

Define versioned workflow DAGs — tasks, prompts, context sources, agents.

02

Run

Execute against real work. Live status per node. Partial re-runs.

03

Observe

Per-node: exact prompt, resolved context, RAG data, model, output.

04

Diagnose

Isolate the failing node. Inspect its full input surface. Experiment.

05

Improve

Run evals against a baseline. Promote what works. Prove it got better.

Platform capabilities

Everything the loop needs.

Workflow Orchestration

Define versioned DAGs of agentic tasks. Author once, run repeatedly, with live per-node status.

Per-node Observability

Every task captures its exact input surface — prompt, context, RAG data, upstream outputs.

Isolated Re-run

Re-run any single task without restarting the pipeline. No waiting for passing nodes.

Eval & Feedback Loop

Attach measurable checks to any task. Quality scores feed back into workflow definitions.

Provider-agnostic

No vendor name in the domain layer. Swap models and providers without touching your DAG.

Theme System

Domain workflows packaged as themes. SDLC is flagship. Extend to any domain without touching core.

Ready to close the loop?

Govern your agentic pipelines.
Improve them scientifically.

Internal-first. Provider-agnostic. Built to improve.

Sign in to your workspace