Hey Jian and the Fractals community 👋
I've been digging into Fractals and think there's a real gap worth discussing: cost visibility and budget control for recursive task decomposition.
The problem is structurally unique to systems like Fractals. A single high-level task with maxDepth=4 can spawn dozens to hundreds of leaf tasks — each making its own OpenAI planning calls and Claude CLI / Codex CLI execution calls. Users have no idea what a task tree will cost until the API bill arrives. That's not a minor inconvenience; it's a blocker for teams that want to use Fractals on anything beyond personal experimentation.
A few specific scenarios where this bites:
- Runaway decompositions — the LLM judges a task to need 40 subtasks instead of 5. No one catches it until post-execution.
- No cost signal at the Review step — users inspect the plan tree before committing to execution, but currently there's no "this will cost approximately $X" estimate to inform that decision.
- No per-branch attribution — if a maxDepth=3 tree costs $18, which branch consumed 80% of that? Was it planning or execution?
This is actually where a recent paper — AgentOS: From Application Silos to a Natural Language-Driven Data Ecosystem — is relevant. It argues that the "Agent Kernel" needs to treat token budget allocation as a first-class OS-level concern across concurrent agent threads. Fractals is one of the clearest real-world expressions of exactly that problem.
I've been building AgentCost — an open-source cost governance platform — that I think maps cleanly onto this. A few integration approaches, from lightest to deepest:
Level 1 — Gateway proxy (zero Fractals code changes):
Point Fractals' OPENAI_API_KEY at the AgentCost gateway instead of OpenAI directly. Every planning-phase call gets captured, attributed, and tracked automatically with no changes to the codebase.
Level 2 — Webhook after leaf execution:
A small addition to executor.ts — POST task metadata (task ID, depth, lineage, executor type) to an AgentCost endpoint after each leaf completes. AgentCost correlates this with LLM usage to produce a per-tree, per-branch cost breakdown.
Level 3 — Cost estimate at Review:
Before the user clicks Execute, call AgentCost's pre-call estimation API with the leaf count and task types. The Review UI shows "Estimated cost: $4.20 – $7.80" alongside the plan tree. Users can set a budget ceiling before committing.
I'd love to hear how people are currently handling this — or whether this is even a pain point for early Fractals users. Happy to build a Level 1 proof-of-concept and share a screenshot of the per-tree cost breakdown in AgentCost's dashboard if that would be useful.
— Vicky
AgentCost | agentcost.in | Demo
Hey Jian and the Fractals community 👋
I've been digging into Fractals and think there's a real gap worth discussing: cost visibility and budget control for recursive task decomposition.
The problem is structurally unique to systems like Fractals. A single high-level task with maxDepth=4 can spawn dozens to hundreds of leaf tasks — each making its own OpenAI planning calls and Claude CLI / Codex CLI execution calls. Users have no idea what a task tree will cost until the API bill arrives. That's not a minor inconvenience; it's a blocker for teams that want to use Fractals on anything beyond personal experimentation.
A few specific scenarios where this bites:
This is actually where a recent paper — AgentOS: From Application Silos to a Natural Language-Driven Data Ecosystem — is relevant. It argues that the "Agent Kernel" needs to treat token budget allocation as a first-class OS-level concern across concurrent agent threads. Fractals is one of the clearest real-world expressions of exactly that problem.
I've been building AgentCost — an open-source cost governance platform — that I think maps cleanly onto this. A few integration approaches, from lightest to deepest:
Level 1 — Gateway proxy (zero Fractals code changes):
Point Fractals' OPENAI_API_KEY at the AgentCost gateway instead of OpenAI directly. Every planning-phase call gets captured, attributed, and tracked automatically with no changes to the codebase.
Level 2 — Webhook after leaf execution:
A small addition to executor.ts — POST task metadata (task ID, depth, lineage, executor type) to an AgentCost endpoint after each leaf completes. AgentCost correlates this with LLM usage to produce a per-tree, per-branch cost breakdown.
Level 3 — Cost estimate at Review:
Before the user clicks Execute, call AgentCost's pre-call estimation API with the leaf count and task types. The Review UI shows "Estimated cost: $4.20 – $7.80" alongside the plan tree. Users can set a budget ceiling before committing.
I'd love to hear how people are currently handling this — or whether this is even a pain point for early Fractals users. Happy to build a Level 1 proof-of-concept and share a screenshot of the per-tree cost breakdown in AgentCost's dashboard if that would be useful.
— Vicky
AgentCost | agentcost.in | Demo