Pillar · Agentic Coding

Agentic Coding: The Loop Is The Product.

Autocomplete writes the next line. An agentic tool owns the task: plan, act, verify, remember. Here is how Anvaya implements each stage, and what is measured versus designed.

Direct answer

Agentic coding means a tool runs its own plan-act-verify loop with real file, shell and search tools, instead of completing a single suggestion. Anvaya CLI is a local-first agentic coding tool: a single static Rust binary with 21 native tools, 4 subagent profiles, a pooled language-server layer, a repomap for structural navigation, and a sandbox policy enforced at the tool layer. It runs the loop headless for CI (--no-tui, exit codes 0/1/2) or read-only for planning (--plan-only). What makes it distinct is the memory layer: turn outcomes can become typed nodes in a local graph that decays by type, checks itself against git drift, and calibrates utility from use. Phase 1 of our open benchmark program measured the harness at 25.6 MB median own-process RSS across 238 runs with 228/238 externally verified completions.

The Loop

Six Stages, Sourced.

Every stage below corresponds to shipped code; the only staged item is in-session plan stepping, and it is labeled.

01Plan firstAnvaya's --plan-only mode is read-only by construction: write and exec tool calls are blocked at dispatch, so exploration cannot mutate the tree. In-session plan stepping is on the roadmap, not shipped.
02Act with real tools21 native tools compiled into the binary, read/edit/patch, glob/grep, run_command and run_script, a pooled LSP, repomap PageRank, web fetch/search, and a mind tool for recall and record. No plugin layer.
03Verify objectivelyTests, builds and external checkers are first-class: the benchmark program treats a passing verifier as the only definition of completion, and the same discipline applies in normal use.
04Isolate the messSubagents (explore, general, plan, build) run in fresh windows and return summaries; up to four run concurrently; a 30-second stale heartbeat surfaces STALLED in the TUI instead of hanging silently.
05Constrain blast radiusA declarative sandbox policy enforces read/write/exec/network scopes at the tool layer. Profiles: project, computer, strict. --yolo bypasses approval but does not widen confinement.
06Remember outcomesEvery turn can feed the Mind graph: typed decisions, patterns and fixes that decay by type, drift-check against git, and carry a Beta-Bernoulli utility score from actual use.

Tool Shapes

Autocomplete, Assistant, Agent.

Same editor, three different contracts. Agentic tools trade speed-of-keystroke for scope and autonomy.

DimensionAutocompleteChat assistantAgentic harness
Unit of workNext editOne questionA task to completion
ScopeCurrent filePasted contextRepo + commands + tests
ToolsNoneRead-only at most21 native tools, sandboxed
VerificationNoneNoneRuns tests/builds, checks results
Across sessionsStatelessStatelessOptional persistent memory graph
CI / headlessNoNoYes, --no-tui, exit codes 0/1/2

Questions

Asked About Agentic Coding.

Q

What is an agentic coding tool?

A coding tool that runs a loop (plan, call tools, observe results, verify, repeat) instead of answering a single prompt. Autocomplete predicts the next edit; a chat assistant explains code; an agentic harness reads files, edits them, runs commands and checks its own work across many rounds under budgets and approvals.

Q

How is agentic coding different from autocomplete?

Scope and verification. Autocomplete is inline, single-file and stateless. An agent holds a task across files, executes commands, reacts to failures and can run headless in CI. The cost is stronger guardrails: approvals, sandbox policy, round limits and context budgets.

Q

Do agentic coding tools need cloud models?

No. Anvaya runs a full loop against local Ollama models with no account or key, and can escalate individual tasks to Anthropic or OpenAI-compatible endpoints. Local models cover everyday edits; frontier models still lead the hardest long-horizon refactors.

Q

What makes Anvaya's loop different?

Two things: measured economy and memory. The Phase 1 benchmark measured 25.6 MB median harness RSS with 228/238 externally verified completions across 238 runs; Mind adds a persistent graph so later sessions can retrieve typed experience instead of re-reading the same files.

Q

Can I run an agentic loop in CI?

Yes, anv --no-tui runs headless with pipeable I/O and exit codes 0 (success), 1 (error), 2 (approval denied). --plan-only gives a read-only planning pass; the sandbox policy and the monitor tool handle long-running processes.

Run Agents That Fit On Your Laptop.

25.6 MB median RSS. 25 agents ran in parallel on a Core 2 Duo with 4 GB RAM. Hundreds on your machine. Zero cloud required on the Ollama path.

Get Started →Read the FAQ

Requires Rust/cargo to build from source. Linux and macOS today, Windows not yet supported. Pre-1.0, public beta. Pricing TBD.