TTacsmith
AI agent platform

Precision AI agents that plan, adapt and control in real time.

Tacsmith builds AI agents for strategic reasoning, physics-based control and adaptive behaviour. Each agent is modular and configurable, and you can run it on our servers or plug it into your own game, simulator or tool.

Read the API docs See plans
Completed projects

Two finished agents, both in action today

Two very different problems: one needs deep, discrete strategic search, the other needs fast, continuous control.

LiveProprietary

Chess AI: strategic reasoning

A chess engine that teaches itself by self-play. Neural networks trained in PyTorch guide a C++ alpha-beta search. The search knows material, finds forced mates exactly and never picks a move proven to lose.

~2600top Elo estimate on a Stockfish ladder
30+trained bots to play against
100%mate-in-one found when available

Elo is estimated from ladder matches at 0.1 s a move, accurate to within about ±60.

Use it through the API →
CompleteReinforcement learning

Racing agent: physics-based control

A driving agent that learns high-speed control by trial and error. It reads speed, tilt, wheel contact, the distances to walls and the shape of the road ahead, and it is rewarded for progress along the racing line.

30 msdecision interval at 400+ km/h
270sensor inputs per decision
27steering and throttle actions

A quantile Q-network with noisy-layer exploration, n-step returns and prioritised replay. One model trains on many tracks at once so it can handle tracks it has never seen.

Watch it drive →
Evidence

See it drive

The racing agent's record run on E02, a 22.6 km endurance track, which it finished in 4'24"47. Nobody is touching the controls: the agent makes every steering, throttle and brake decision itself. The overlay at the top left shows its inputs as it drives.

The first 37 seconds of the 4'24"47 record run (8 MB).

From zero to a 4'24" lap in about 5 hours

The same agent learning E02 from scratch, taken from its training log. Each attempt is one run of the track.

Training timeBest result
10 min38% of the track, no finish
30 min97.5% of the track
1 hFirst finish: 5'34"50
2 h4'37"32
3 h4'31"30
4 h4'29"71
5.3 h4'24"47, attempt 5,527

That's 70 seconds faster than its first finish. With no human demonstrations, the agent learned it purely from trial and error.

Platform

Built as modules, not one-off bots

Every agent is made from the same building blocks, so you choose the behaviour, where it runs and how you measure it.

Behaviour profiles

Tune aggression, risk, style and difficulty without retraining. Save profiles and reuse them across sessions and titles.

Server-side execution

Agents run on our infrastructure behind a simple API. Your client sends state and receives actions back.

Analytics

Track results, ratings, decision timing and behaviour over time, so you can see how agents perform against your players.

Simulation

Run batches of agent-vs-agent games or simulated episodes faster than real time to test balance before release.

For game studios

Gameplay AI that improves the player experience and saves development time

The same techniques behind the chess and racing agents carry over to the AI systems studios need most. Strategic opponents are available now through the chess AI. For the rest, Tacsmith is working on new bots. See In Development for each one's status.

Strategic AI opponents

Opponents that plan ahead, adapt to how each player plays and give a real challenge without feeling unfair.

NPC behaviour systems

Characters with believable routines, reactions and personalities that make a world feel alive.

Pathfinding and navigation

Units that move intelligently through complex environments, avoid obstacles and choose the best routes.

Resource and worker automation

Agents that gather, build, craft and optimise production chains in management and colony games, with minimal scripting.

Combat decision making

Targeting, positioning, threat evaluation and tactical choices for combat-focused titles.

Automated testing bots

Bots that play through levels, stress-test systems and find bugs far faster than manual QA.

Integration

Send in the state, get back the next action

Every agent uses the same request-and-response pattern over HTTPS. You can start with a single request, then add behaviour profiles, sessions and analytics as your integration grows.

Integration guide 
curl https://api.tacsmith.com/v1/chess/move \
  -H "Authorization: Bearer $TACSMITH_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "fen": "rnbqkbnr/pppppppp/8/8/4P3/8/PPPP1PPP/RNBQKBNR b KQkq - 0 1",
    "bot": "ai-latest",
    "profile": "club-player",
    "limits": { "time_ms": 200 }
  }'

{ "move": "c7c5", "eval": 0.04, "depth": 14, "time_ms": 197 }
Closed source. Tacsmith's agents and training systems are proprietary and are available only through the hosted API. Paid plans include commercial use: see licensing.