Research in public.
Where we're honest about the frontier: the problems we're chasing and the benefit we're after. We publish the direction, and keep the mechanism in the lab until it ships.
// direction published · mechanism kept · nothing here is a shipped feature
The experiments
Four directions we're chasing
Each one as its own section — why it matters, what we're exploring, and where it stands. Jump with the bar, or scroll the whole story. We publish the direction; the mechanism stays in the lab until it ships.
Agent roles
Why a small team of scoped agents beats one generalist trying to do everything.
each role hands a reviewed artifact to the next
A single generalist agent has to be an expert at product, architecture and delivery all at once. Splitting responsibility — the way a strong organization does — makes each proposal more grounded, and far easier for a human to review.
Where it stands. The direction shows up in Orchestration, which is Planned. This is research, not a shipped feature.
What we're exploring
- H1Scoping an agent's responsibility so it stays in its lane
- H2How specialized roles hand work off to one another
- H3Where a human's judgment sits between them
Run memory
Runs that start from what your organization already decided — instead of a blank slate.
Every run today re-derives context that already exists somewhere. Memory that carries decisions, constraints and rejected alternatives across runs is what makes the next one faster, and more consistent with the last.
Where it stands. TaskForce Memory ships today in Beta. The deeper run-to-run learning is still research.
What we're exploring
Model choice
The right model for each step — local or hosted — instead of one model for everything.
Different steps have different needs: some want a fast local model, some a stronger hosted one, some must stay on your own hardware. Choosing per step keeps cost, latency and privacy under your control.
Where it stands. Running on your models ships today — local through Ollama, or hosted. Automatic per-step routing is research.
- Draft specLocalon your hardware
- Generate codeHostedstronger model
- Review & lintLocalfast · low cost
routed by need — privacy, strength, cost — not one model for all
What we're exploring
Learning from reviews
What you edit, reject or send back should make the next run better.
each correction is a signal — the next proposal lands closer
Every override is a signal. A system that learns from what humans correct — without training on your private data — could calibrate its own proposals over time, and eventually tell you how reliable they are.
Where it stands. This is the Prediction & calibration direction — Planned, not shipped. It's the honest edge of what we're building.
What we're exploring
- H1Turning overrides into signal
- H2Calibrating proposals without training on your data
- H3Measuring whether a decision was actually right
The system
The pieces are one loop
The experiments aren't separate features — they're steps of one loop TaskForce reasons over. Intent enters, runs against what your organisation already decided, and every decision feeds back into memory.
↺ the decision feeds back into memory — the loop closes
From the lab to the product
What has already graduated
When something works, it leaves Labs and ships with a maturity label. These started as questions here — they're in the product today.
Follow where this goes
The roadmap is public — hold us to it. And start with what already ships today.