See the work, the cost, and the behavior.

Orchestration is only half of it. The other half is knowing what your agents did, what they cost, and how they behaved — across every run, every workspace, every product. This is the deepest-built part of Fragua, and the reason a team can run many agents without losing the thread.

The three axes

Watch the work as the agent produces it.

Every durable event from each turn lands in the database one row at a time, and the browser sees a live timeline — streaming, not polling. The server pushes each turn over Turbo Streams as it happens; the browser never asks.

The log is durable, not chatty: assistant text, mid-interview questions, tool results, terminal artifacts — persisted as the full firehose, filtered at render time. A session id is pre-generated on each run, so the work survives a closed browser or a crash. Come back tomorrow; the log is still complete.

Almighty monolith · run 14 · live

19s · $0.18

SYSTEM 20:01:48

Run started · Almighty monolith · claude-sonnet-4-6
TOOL CALL 20:01:50

Read app/models/invoice.rb
TOOL RESULT 20:01:51

ok · enum :status draft/sent/paid
TOOL CALL 20:01:54

Edit app/models/invoice.rb
TOOL RESULT 20:01:55

ok · 5 lines added
TOOL CALL 20:01:58

Bash "bin/rails test test/models/invoice_test.rb"
TOOL RESULT 20:02:05

ok · 14 runs, 28 assertions, 0 failures
SYSTEM 20:02:07

Opened pull request · feat/invoice-status-enum

A Fragua agent run · illustrative

This month

$4.32

8.1k in · 24.3k out · 4.0M cache

Daily spend last 30 days

May 12 Jun 10

Monthly cost · illustrative

By phase

Research $0.00 0%

Plan $0.18 4%

Spec $0.61 14%

Execution $3.53 82%

Cost by phase · illustrative

Effort · one feature

Cost: $1.18
Time: about 2 hours
Tasks: 13/13
Builds delivered: 1
Tool actions: 193

Effort · one feature · illustrative

Every token, rolled up and accounted for.

Fragua parses the usage events from every run into their own table — one row per turn, with the model and token counts. Then it rolls them up so finance knows where every dollar went: per run, per workspace, per account, per month, and per phase — across research, plan, foundation, spec, and execution.

This is metering, not gating. We never proxy your AI calls — tokens are billed directly to your provider, on your key, so we can't see your bill or charge you for usage. The model and reasoning effort for each phase are set in one place (config/agent_models.yml) and recorded on every turn, so a cost is always traceable to the decision that ran it.

The record of how each agent behaved.

The durable event log is more than a transcript — it's the record of how each agent behaved. What it touched, what it ran on, why it stopped. Every run carries the same fields, so a stalled or failed run is never a mystery.

A heartbeat watchdog catches stalled runs and restarts cascade cleanly without losing the log. Discarded and restarted runs still count toward roll-ups, so a budget can't be reset by starting over.

Run history

10 runs

Run #10 completed
1 day

Daniela Ruíz Execution 42m 18s
Run #9 failed
1 day

Daniela Ruíz Shape spec 31m 49s

stalled: no activity in the heartbeat window
Run #8 cancelled
2 days

Mateo Fuentes Execution 6m 02s
Run #7 completed
3 days

Daniela Ruíz Shape spec 8m 25s
Run #6 completed
7 days

Mateo Fuentes Execution 57m 9s
Run #5 completed
7 days

Daniela Ruíz Shape spec 12m 20s

Run history · illustrative

Run agents you can actually see.

Fragua is in private beta. Join the waitlist and tell us about the work.

Request access