F-00 MISSION CONTROL FOR MULTI-HARNESS OPERATORS
Run the operation,
not the agents.
Claude Code, Buzz, Codex — the harnesses run sessions. You operate a portfolio. Forge is being rebuilt as the layer above. The target: one outcome ledger and one decision queue across every agent and venture — what ran, what it cost, what actually shipped, and what needs your GO.
WEEK 32 · HAND-RECONSTRUCTED ESTIMATE
F-01 THE SHIFT
Every season, the harnesses eat another layer of the agent stack. Good. Let them.
Orchestration, memory, model routing, workflows — the things agent platforms promised in March are native harness features by August. Building there means racing the labs. You lose.
What no single harness gives you across the fleet is the view from above the sessions: the money, the fleet-wide record, the go/no-go call. A mixed fleet across several ventures, and nobody can say what this week cost or what actually shipped.
That layer — the loop above the runtime — is Forge.
F-02 STILL OPEN
Four jobs between
the sessions
EVERY WORKFLOW BELOW CAME FROM A REAL OPERATING FAILURE.
ARTIFACTS: TARGET STATE · DRAWN FROM LIVE DOGFOOD
01
Budgets that bite
Hard caps per agent, per venture, per month. At the cap, Forge will pause supported work and open a decision — extend, reroute, or stop — rather than merely returning a 429. The target: no token-bill surprise at invoice time.
02
One fleet ledger
Every agent, every venture, one page: what ran, what it cost, what shipped. Forge will answer the Friday question continuously instead of reconstructing it from scattered transcripts.
03
Autonomy with a gate
For work launched through supported integrations and hooks, Forge will put a human gate in front of irreversible actions. Deploys, spends, deletes, and publishes queue for your GO. Forge owns the decision queue — not arbitrary runtime interception.
04
Evidence, not vibes
A Forge "done" record will link to proof: the diff, test run, and deploy log — or the sources, data snapshot, and publish receipt. Independent verification is shaped to the job and kept past the session.
F-03 PROOF OF NEED
A week of our own fleet, by hand.
| Agent | Role | Shipped this week | Sessions | Tokens | Cost | State |
|---|---|---|---|---|---|---|
| fizz | lead engineer | antemass rebuild → preview; signal scroll-world | 14 | 21.3M | $64.20 | GATED |
| mason | project engineer | meridian-pier.vercel.app live | 9 | 12.8M | $37.80 | IDLE |
| honey | chief of staff | 3 pSEO pilots; this page | 11 | 8.9M | $26.90 | ACTIVE |
| bumble | researcher | GSC keyword re-mine, queued | 3 | 2.1M | $6.10 | SCHEDULED |
| WEEK 32 TOTAL | 37 | 45.1M | $135.00 | |||
This table was reconstructed by hand from four agents' scattered transcripts. That it had to be is exactly the problem Forge exists to end.
F-04 DOCTRINE
Own the loop.
Rent the runtime.
The harness runs the session. Forge runs the operation.
Labs will keep improving the runtime. That's their race — rent it.
Money, record, and judgment stay yours. That's the loop — own it.
F-05 FLIGHT STATUS
Being refit in the open.
Forge began as a 136,000-line monorepo, including roughly 83,000 lines of runtime, built to replace the harness. The harness won. We audited the useful pieces as a quarry and locked a fresh thin-service architecture above it — then put product code behind an Aug 18 dogfood gate.
The bar for early access: Forge runs mission control for our own mixed fleet across several ventures for a full month, and we stop reconstructing ledgers by hand. Until then, PRE-FLIGHT is literal.
LAUNCH CHECKLIST
- ■quarry audited; thin-service architecture locked
- □budgets + hard-stops on live fleet
- □fleet ledger, updated continuously
- □gates on every supported irreversible action
- □one month, zero hand-built ledgers
- □open the doors